Vous êtes sur la page 1sur 86

When All You Have is a Banhammer: The Social and

Communicative Work of Volunteer Moderators


by
Claudia Lo
B.A., Swarthmore College (2016)
Submitted to the Department of Comparative Media Studies
in partial fulfillment of the requirements for the degree of
Master of Science in Comparative Media Studies
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
June 2018
© Claudia Lo, MMXVIII. All rights reserved.
The author hereby grants to MIT permission to reproduce and to distribute
publicly paper and electronic copies of this thesis document in whole or in
part in any medium now known or hereafter created.

Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Department of Comparative Media Studies
May 11, 2018

Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
T. L. Taylor
Professor of Comparative Media Studies
Thesis Supervisor

Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Heather Hendershot
Professor of Comparative Media Studies, Director of Graduate Studies
2
When All You Have is a Banhammer: The Social and Communicative
Work of Volunteer Moderators
by
Claudia Lo

Submitted to the Department of Comparative Media Studies


on May 11, 2018, in partial fulfillment of the
requirements for the degree of
Master of Science in Comparative Media Studies

Abstract
The popular understanding of moderation online is that moderation is inherently reactive,
where moderators see and then react to content generated by users, typically by removing it;
in order to understand the work already being performed by moderators, we need to expand
our understanding of what that work entails. Drawing upon interviews, participant obser-
vation, and my own experiences as a volunteer community moderator on Reddit, I propose
that a significant portion of work performed by volunteer moderators is social and com-
municative in nature. Even the chosen case studies of large-scale esports events on Twitch,
where the most visible and intense tasks given to volunteer moderators consists of reacting
and removing user-generated chat messages, exposes faults in the reactive model of mod-
eration. A better appreciation of the full scope of moderation work will be vital in guiding
future research, design, and development efforts in this field.

Thesis Supervisor: T. L. Taylor


Title: Professor of Comparative Media Studies

3
4
Acknowledgments
To T. L. Taylor, for her unwavering support for both my thesis-related and non-academic
endeavours; to Tarleton Gillespie, my reader, for his generosity and thoughtful insight; to
Kat Lo, fellow partner-in-academic-crime; to Shannon, CMS’s very own chocolate-bearing
problem-solving wizard extraordinaire. To my cohort, with whom I have endured this pro-
cess and to whom I am indebted to for so much.
To the ESL moderation team who walked me through the baby steps of Twitch moder-
ation with true grace; to DoctorWigglez, whose help left this thesis far richer.
To a certain verdant waterfowl, who taught me everything I know about moderation;
to my moderation team on reddit (past, present and future) from whom I have learned so
much; to the Euphoria regulars who provided me with feedback, support, and an uncanny
ability to help me work out what I was saying better than I did myself; to the denizens of the
Crate & Crowbar, menacing with spikes of pure wit and adorned with puns of the highest
calibre, without which I would be short, amongst other things, a title; to the Cool Ghosts
of the internet high-fiving me through the wee dark hours of the night as I made my way
through the process.
To Erin, for everything:

My heartfelt thanks and deepest praise to you,


The seed of this was not of mine alone.
Without your constant guidance to turn to,
This thesis, stunted, would never have grown.
Yet with your care came blossoming of prose,
In ink it flowered and now lays in repose.

5
6
Contents

1 Introduction 11
1.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2 Models of Moderation 21
2.1 The reactive model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 The lasting power of the reactive model . . . . . . . . . . . . . . . . . . . . 26
2.3 Counter-model: the proactive model . . . . . . . . . . . . . . . . . . . . . 28

3 Moderation on Twitch 31
3.1 Moderation tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Running an event . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2.1 Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2.2 During the Event . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2.3 Cleanup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3 The reality of event moderation . . . . . . . . . . . . . . . . . . . . . . . . 51

4 The Social World of Moderation 57


4.1 The life of a moderator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2 The relationship between moderators and Twitch chat . . . . . . . . . . . . 64
4.3 What is good moderation? . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3.1 Rogue moderators, online security, and handling threats . . . . . . 71
4.3.2 Badge-hunters and proper moderation values . . . . . . . . . . . . 74

7
5 Moderation Futures 79
5.1 Twitch, esports, and event moderation . . . . . . . . . . . . . . . . . . . . 79
5.2 Transparency, accountability, and reporting . . . . . . . . . . . . . . . . . 81

8
List of Figures

2-1 A diagram of the reactive model of moderation. . . . . . . . . . . . . . . . 22

3-1 A screenshot of Twitch. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32


3-2 An example of Logviewer, showing multiple user chat histories, with mod-
erator comments on a user. . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3-3 FrankerFaceZ’s moderation card. . . . . . . . . . . . . . . . . . . . . . . . 38
3-4 An example of a moderator’s triple-monitor setup. . . . . . . . . . . . . . . 55

4-1 Some examples of popular mod-related spam. . . . . . . . . . . . . . . . . 68


4-2 Global Twitch face emotes often used in offensive messages. From left to
right: TriHard, cmonBruh, HotPokket, Anele . . . . . . . . . . . . . . . . . 76

9
10
Chapter 1

Introduction

One night, as I was preparing to leave for class, I got a message notification from the cha-
troom that I help moderate. A user, posting in a specific area of the chatroom meant for
LGBTQ users, asked if they were allowed to ask not-safe-for-work (NSFW) questions. This,
in turn, sparked off pages of fast-moving speculation: what qualified as NSFW? How would
the moderators respond? What policy had Discord, the platform provider, set out? In an at-
tempt to assuage these fears, I ended up creating an on-the-fly preliminary policy regarding
the posting of potential explicit content, while trying to deal with the reasonable concern
that this LGBTQ chat would be cordoned off as an 18+ space, without going against Dis-
cord’s Terms of Service. All while swiping wildly on my phone keyboard, trying not to fall
down the stairs leading into the subway station.
The next day, my mod team and I evaluated the impact of this decision for both our
Discord chatroom and our related community forum on Reddit. In the parlance of my fel-
low volunteer moderators, neither the banhammer nor the broom was needed: that is to
say, no one needed to be banned, and no content needed to be swept away. Nothing was
removed, so by the popular standards of online moderation, no moderation had happened.
Yet this kind of decision-making forms the most significant and challenging aspect of my
work as an online volunteer community moderator. Popular perceptions of online modera-
tion work, both volunteer and commercial, portray it quite differently. Disproportionately,
the discourse surrounding moderation, and related topics of online abuse, harassment, and
trolling, centers on a small set of actions that I do as a moderator.

11
In my time online, as a participant and as a moderator on various platforms, the presence
of human community moderators was common and everyday knowledge. Yet the discourse
surrounding online moderation, particularly as it pertains to online social media platforms,
takes quite a different route, instead hiding and suppressing the very presence of moderation
as much as possible, and associating it less with human actors than to algorithmic processes
and platform design.
The language of affordances here, drawing from Latour’s actor–network theory, will be
particularly helpful to figure out how the peculiarities of different online platforms shape
the nature of the communities that they culture, as well as the forms of moderation and
regulation that take place upon them. Online content moderation as it occurs on different
online platforms has been the topic of increasing academic interest. I will borrow Gillespie
(2018)’s definition of platforms for this thesis:

For my purposes, platforms are: online sites and services that

a) host, organize, and circulate users’ shared content or social interactions


for them,

b) without having produced or commissioned (the bulk of) that content,

c) built on an infrastructure, beneath that circulation of information, for


processing data for customer service, advertising, and profit.

I am aware that ‘platform’ is a term that is thrown around quite liberally when discussing
this subject. Hence, I want to distinguish between forms of moderation conducted on plat-
forms, but by different stakeholders. The majority of work on moderation has centered mod-
eration performed by platform operators, on the platforms that they run. Though it sounds
tautological, this distinction is important: specifically I wish to divorce the idea that the
reality of carrying out moderation on a platform always primarily rests on the platform op-
erator. Such moderation work—that is, moderation performed by platform operators—has
been variously described as intermediary governance (Gasser and Schulz, 2015), as the gov-
ernors of online speech (Klonick, 2017), and as the self-defense of a semicommons (Grim-
melmann, 2015). At this point, though, I would like to point out that such descriptions of

12
moderation render invisible the people who carry it out: those who design the boundaries
of these places, who create the different tactics and policies that constitute this governing
and regulatory work, and who ultimately carry them out.

Online moderation in all its forms has enormous impact on the experiences of millions,
and potentially even more, as online social spaces proliferate. Yet even as current events
make apparent the need for moderation in online spaces, we are, generally speaking, go-
ing into this practically blind. The public appetite for platforms to regulate users grows day
by day, and yet we are unclear as to what it is we want, and how it should be done. More-
over, public discourse tends to place the responsibility of regulation upon platform operators
alone; while for various political, rhetorical, practical and moral concerns, this may make
sense, I fear that defining this argument with platform operators as the only group tasked
with moderation blinds us to the efforts of community-led moderation.

I would propose two basic types of moderation: decontextualized moderation, and con-
textualized moderation. Decontextualized moderation is characterized by the fact that those
who conduct this work are alienated from the community of users whom they are expected
to moderate. Commercial content moderation as described by Roberts (2012) is one defin-
ing example: these moderators generally formally employed by the same company that runs
the platform(s) upon which they moderate, but are distanced in multiple ways: geograph-
ically, being contracted workers far removed from the countries where the company itself
may operate; technologically, by providing these workers with a controlled portal that does
not allow them to seek out additional contextual information; and socially, by removing
them from the platform itself. Additionally, the focus of commercial content moderation
tends to be the content on the platform, rather than user behaviours or norms, and the ac-
tions that can be undertaken by these moderators is accordingly limited to either removing
it, or leaving it alone. Decontextualized moderation would also extend to non-human agents
that perform moderation work: the “algorithmic configurations” (Humphreys, 2013) that
promote, suppress, and otherwise shape the contours of online social platforms. Examples
of commercial content moderation might include an Amazon Mechanical Turk worker paid
a few cents per image to decide whether or not an image is impermissible, either perform-
ing such moderation work directly or providing the human judgment necessary to train AI

13
other machine learning algorithms to eventually perform this work.
Contextualized moderation, on the other hand, is generally performed by people drawn
from the same communities that they then moderate. This may be paid, as in community
management positions, or unpaid volunteer labor. They work with a community that is ex-
pected to persist, forming expectations and norms that will impact moderation work. There
are many striking similarities between professional and amateur contextualized moderation
work. In brief, both “sit between developers and game players in the production network but
are an important element in maintaining capital flows” (Kerr and Kelleher, 2015) although
the presence or absence of an explicit contract will affect the relations of power as well as
the responsibilities and implications of the role. Different platforms will additionally have
different affordances which further impact the work of community moderators. The focus
of their work is on the well-being, conduct, goals, and values of a given community, which
encompasses any content created or found within it.
It should be noted that these are not mutually exclusive forms of moderation. Indeed, I
would be surprised to see a platform that employed only one or the other regardless of their
rhetoric. Community managers and moderators may employ forms of decontextualized
moderation to do their work; for example, they may employ contracted workers in order to
train automated tools, or use algorithmic methods to implement moderation policies that
they devise. Conversely, the outcomes of decontextualized moderation may impact the work
of contextual moderators; the rollout of a platform-wide moderation algorithm affects the
work that the embedded moderators of that platform will then perform.
Additionally, these different types have different strengths. Most notably, contextualized
moderation relies in some part on understanding the cultural and social norms and values of
a community, thus presupposing the existence of a community in the first place. While the
term itself is often thrown around by platform operators themselves, referring to ‘a Twitter
community’ or ‘a Facebook community’ or ‘a Reddit community’, it is safe to say that such a
platform-wide community exists only in the abstract. I turn to Preece (2000) for a working
definition of community:

An online community consists of:

14
1. People, who interact socially as they strive to satisfy their own needs or
perform special roles, such as leading or moderating.

2. A shared purpose, such as an interest, need, information exchange, or


service that provides a reason for the community.

3. Policies, in the form of tacit assumptions, rituals, protocols, rules, and


laws that guide people’s interactions.

4. Computer systems, to support and mediate social interaction and facili-


tate a sense of togetherness.

That is to say, these elements shape and direct moderation, and moreover that any given
platform supports not one but a myriad sub-communities, with no guarantee that any one of
their four constituent elements respect the boundaries of platforms on which they operate.
As we continue on to look at moderation of communities, it is important to note that these
elements are at once keenly felt by their members, yet also flexible, ambiguous, and fuzzy
with respect to their borders. Thus, even as community moderators react to nature of said
community as it pertains to their work, there is a degree of flexibility and deep cultural
awareness at play.
There is a basic, formal distinction exists between different classes of online commu-
nity members on the Internet: simply put, volunteer moderators are users given limited,
but disproportionately powerful, permissions to affect what other users can or cannot see
and do on the platform. This puts them uncomfortably between regular users, with no such
special permissions, and platform operators or administrators, who have full permissions
to affect the running of the platform; in the most extreme case this constitutes access to the
literal on/off switch for the servers themselves. In a platform where some level of regulatory
power is distributed, for example manipulating a comment ‘score’ that in turn affects the
comment’s discoverability, one would expect a moderator-user to have permissions above
and beyond this base level. On Reddit, where every user has the ability to manipulate com-
ment score through a single up- or down-vote, per comment, a moderator can remove it,
rendering its score moot; this would be an example of that ‘disproportionately powerful’
ability to affect content for other users. However, moderators on Reddit cannot shut down

15
other communities, or ban users from the platform itself; those permissions are only granted
to administrators, who are employees of Reddit. In that sense, a moderator both has editing
permissions beyond that of a regular user, but below that of an administrator or platform
operator.

These volunteer moderators have been variously portrayed as exploited by capital as part
of the “free labor” that keeps the Internet running smoothly (Terranova, 2000), or analyzed
through more of a co-creative lens as laid out by Jenkins (2008). However, this model of
user-generated moderation, distinct from various forms of commercial content moderation
(Roberts, 2012), has been complicated in recent years. Rather than understand these users
merely as exploited users, or as equal creative partners, volunteer moderators work within
an alternate social structure of values, motivations and norms that is influenced and shaped
by capital and existing structures of power, yet does not necessarily respect their boundaries
and strictures. This is a similar complication to that raised by Postigo (2016) in his 2016
analysis of YouTubers and how they generate revenue from the site.

Volunteer moderators and volunteer moderation has been described in many different
ways, as peer-produced oligarchic institutions (Thomas and Round, 2016; Shaw and Hill,
2014), as sociotechnical systems (Niederer and van Dijck, 2010; Geiger and Ribes, 2010), au-
tocratic “mini-fiefdoms” enabled by platform policy (Massanari, 2015), performers of civic
labor (Matias, 2016), moral labor (Kou and Gui, 2017), and as negotiated peer regulation
and surveillance (Kerr et al., 2011). The wide range of these descriptions suggests an equally
broad subject matter: that is to say, moderation, in different spaces for different commu-
nities, may be called upon to perform all manner of roles. Nor do I believe any of these
are necessarily mutually exclusive. Much like the mixed reality of contextualized, decon-
textualized, professional and amateur labor that comprises online moderation, what exactly
moderation is is equally mixed and dynamic. Quite simply, the work of volunteer modera-
tors, even a very narrow subset, is complex enough that we stand to benefit from a broader
picture of that work, to better compliment what work exists. In particular, I want to locate
the human workers in this work.

I am not necessarily proposing a ‘how-to’ guide for either platform operators or volunteer
moderators. In contrast to broader work on online community building, such as Kraut and

16
Resnick’s guide on building online community, I do not want to simplify this down to a
matter of design. Rather, my aim is to refine our understanding of, and perspectives on,
online volunteer moderation. If we think of moderation in terms of what platform operators
are doing, what are we missing out? And if we think of moderation as the removal of content,
what do we render invisible to ourselves merely by having such a narrow definition? It is
valuable to fill out that part of the everyday work that makes online social spaces what they
are, because

I will focus on a particular subset of contextualized moderation, event moderation. This


is moderation of a community that is centered on a specific event, and is therefore limited in
time and scope following the contours of that event. Event moderation may be conducted on
multiple platforms simultaneously, and its audience come together because of the event and
disperse once it ends, though they might rarely form the basis of a longer-term community.
For very well established events that recur on a regular basis, such as a yearly esports tourna-
ment, a community of regular users may also develop, but generally speaking the majority
of the audience are newcomers or brief visitors and therefore once may not expect them to
develop the same kinds of interactions with moderators as more stable communities do.

More specifically, I work with at large-scale esports event moderators on Twitch. These
moderators work for large esports tournaments, which might be expected to draw several
hundred thousand concurrent viewers at their peak. These tournaments generally run for
a few days, over a weekend, and are put on by organizations that work together with game
developers to run tournaments. The games covered by the moderators interviewed included
Valve’s Defence of the Ancients 2 (DOTA 2) and Counter-Strike: Global Offensive (CS:GO);
Riot Games’ League of Legends (LoL); Blizzard’s Hearthstone; and most recently, Bluehole
Studio’s PlayerUnknown’s Battlegrounds (PUBG). Of these games, the two most common
were DOTA 2 and CS:GO, with both drawing huge crowds. Some moderators would attach
themselves to a particular game, moderating based on their existing fan attachment to that
game, but it was not uncommon for them to work for the same event organizer on multiple
concurrent events, whether or not they featured the same game.

Why this narrow focus? My intention is to demonstrate that even for a population of
moderators whose most common action is the removal of content, there are many more im-

17
portant aspects to what they do, and how they conduct it, that has been overlooked. Even
allowing for the most generous fit for our current Furthermore, these overlooked aspects are
not merely complimentary to moderation-by-removal, but integral in guiding their moder-
ation, both in the judgment and in the labour of performing these tasks.

1.1 Methodology

I conducted nine in-depth interviews with Twitch esports moderators with ethical approval
from MIT. All of the interviewees had worked on large-scale events, here defined as recur-
ring semi-regular esports tournaments that were expected to draw concurrent viewer counts
of over 100,000. The largest regular event that some of these moderators covered were the
Majors for CS:GO, including 2017’s record-breaking ELEAGUE Major, which peaked at
over a million concurrent viewers.
I also sat in on two large moderator-only Discord servers, one for ESL moderators and
another more general Twitch moderation server, under a marked “Researcher” account,
with which I solicited these interviews. While I reached out to most of my interviewees, a few
volunteered to be interviewed and would recommend others to me to be interviewed. Their
quotes here have been lightly edited for grammar and to preserve anonymity. The ques-
tions for my interviews were largely based from some preliminary conversations I had with
Twitch moderators, and also drew from my seven years’ experience as a volunteer modera-
tor on some large (over 100,000 subscriber) groups on Reddit. All interviewees were given
a consent form to sign and were allowed to view an advanced copy of this thesis.
I was also granted moderator status on the esl_csgo channel, and engaged in partici-
pant observation, starting with the Intel Challenge tournament on 26 February 2018. This
channel broadcast the Electronic Sports League’s Counter-Strike: Global Offensive events,
and would draw upwards of 90,000 concurrent viewers during live broadcasts. The mod-
eration team were aware of my motivations for joining, and I was expected to perform the
duties of a junior moderator in accordance with their guidelines and existing moderation
precedent.
I used my own regular Twitch account for this, but was given access to the ESL moder-

18
ation guidelines and so changed my setup to fit. This meant that I had to set up two-factor
authentication on Twitch using the Authy app, sit in on their moderator Discord to remain
in contact through the event, and was granted access to Logviewer for that channel. I did not
have access to any bot settings. My primary focus was moderating the newly-added Rooms,
one for each team playing, and I was not focusing on the main stream chat. I also relied
on the help of moderators to understand the different meanings of the various emotes and
more famous memes circulating on Twitch. This is especially needed if trying to read a chat
log, since many of the emote names are in-jokes that have since expanded out, and at any
given time an emote may be used for its surface-level appearance or for the in-joke that it
celebrates.
Twitch event chat, with its fast pace and emphasis on repetition of in-jokes and memes,
can be extremely intimidating at first pass. However, many features of moderator-facing
chat clients, available for free, are also immensely helpful for researchers. It is vital to note
that, as Twitch undergoes constant updates and revisions to its APIs, these third-party tools
are liable to break or suddenly have limited functionality until their developers can support
whatever is the latest version of Twitch chat.
Generally speaking, many of the features that moderator-facing clients, or other third-
party plugins, implement are also extremely useful for researchers trying to get a grasp of the
size, scale, speed and tone of a given channel. For watching chat live, plugins that allow for
pause-on-mouse-hover are invaluable for keeping up with chat, which is offered by Franker-
FaceZ1 or Better Twitch TV, though the former has far more active developer support.
When watching live, other moderator-facing clients I used were 3ventic’s Mod Chat
Client, and CBenni’s MultiTwitch. Multi-Twitch allows one person to see several chats next
to each other at once, while the Mod Chat Client highlights and states which messages have
been removed, and crucially, by whom. However, neither of these programs generate logs,
and therefore are useful only if the researcher is also taking notes during the stream.
To generate logs of Twitch chat, I used both an IRC client, Hexchat, and a custom Twitch
chat client, Chatty, which was designed for moderators. Both create chat logs as .log files

1
As of time of writing, FrankerFaceZ has most of its features disabled as a new version is written for com-
patibility with the latest updates to Twitch chat. It also cannot affect Twitch Rooms.

19
which can be easily converted into a plain text file. Both of these clients can log ban or
timeout messages, and preserve messages which are later deleted. In a plain text format,
emotes are not preserved; instead they are represented by the name of the emote. Chatty also
allows one to view a moderation log and AutoMod settings if the account used to connect
to the channel has moderator status. However, even without moderator status, Chatty is
extremely useful as it allows for keyword or keyphrase highlighting, looking at individual
users’ chat history, and to see charts of viewer count over time.
Twitch chat logs were also downloaded after-the-fact using a Python script, rechat-dl.
This script downloads the JSON file containing recorded chat information, and this was
converted to a CSV file using R, while also stripping out extraneous information. It should be
noted that Twitch currently allows viewers to leave timestamped chat messages on replayed
streams, meaning that this should not be understood as a perfect archival copy of the stream
and stream chat. The stored messages also have some formatting quirks which must be dealt
with; one major one is that deleted lines are preserved only as blank lines. This means it is
possible to see how much of chat was deleted, but it is impossible to guess why, or to see who
performed this action. Another minor issue is that timestamps are saved in epoch time, and
I am currently unsure what Twitch uses for their epoch date.

20
Chapter 2

Models of Moderation

2.1 The reactive model


I call the existing understanding of volunteer moderation work the reactive model. At its
core, this model positions moderators and their work as perpetually reactive, responding
to what users do. It is both a narrative of moderator action, and an ideal. The narrative is
that a user does something, a moderator sees it, and then the moderator either decides to
do something or nothing in response.
The ideal form of moderator action is seamless, in that it should leave no or minimal
trace. For example, if a moderator removes a comment, the remaining trace should not draw
attention to itself, or it should be totally invisible. This is because moderator action is seen
as an exception to the normal user experience. If “nothing” is what moderators normally
do, when moderators do something, it is imperative that the disruption be minimal.
By and large, moderator action is conceived of as the removal of user content, or of users
themselves. There exists a sizeable taxonomy of different forms of moderator actions aimed
at removing content, or otherwise putting up a barrier to its legibility. Aside from total
comment deletion, different platforms offer moderators different tools: for example, the
practice of “devowelling” or “disemvowelling”, where all the vowels in a given comment are
removed, was a popular way to make disruptive comments harder to read (Kraut et al., 2011).
The different forms of banning users are equally diverse. There are bans based on duration
(temporary versus permanent versus “kicking”, which does not stop one from logging in

21
Figure 2-1: A diagram of the reactive model of moderation.

to the space again); there are bans based on identifiers (username bans versus the more
extreme IP ban); and lastly there are bans based on formality (a regular ban that gives the
banned user a notice, versus shadowbans, where any comment made by the banned user is
immediately removed without that user’s knowledge). Automated tools, whether third-party
or built into the platforms themselves, further expedite this process by giving moderators
access to blacklists of words, phrases, or more sophisticated pattern-matching tools. This
automation also revolves largely around the removal of content.
Outside of removal, some platforms are beginning to give moderators ways to promote
content, or otherwise positively interact with content. These may take the form of promot-
ing, distinguishing, or otherwise highlighting comments or users whose behavior or con-
tent exemplifies values held by that particular community. For example, in Sarah Jeong’s
The Internet of Garbage, one suggested “norm-setting” moderation action is “positive rein-
forcement”, the demonstration of good content or behaviours However, these positive inter-
actions are still about reacting to content; a similar practice is suggested by Grimmelmann
(2015). The only difference between such promotions and content removals is that, instead
of obscuring content, positive moderator reactions make selected content easier to find. The
tools available for positive moderator actions are also relatively unsophisticated; in contrast
to the adoption of automated tools for comment or user removal, promotion of content is
still largely done according to a moderator’s human judgement. This reactive model explains

22
why total automation of moderation is seen as a plausible solution to online harassment
and general negativity on a site. If moderation is seen as spotting red flags—words, phrases,
emotes, usernames, avatars or similar—and responding to them, then total automation sac-
rifices little for a lot of potential gain. Under a reactive model, human judgement is only
needed in the most ambiguous of cases, an outlying scenario easily handled by a skeleton
crew. Proper punishment, having already been matched up to an appropriate behavior in
internal moderation policy, can be doled out by the same tool. Automation would provide
far greater reach and pace that human moderators could hope to achieve, at a fraction of the
costs.

The overwhelming perception of moderation work is that removal of content lies at its
core. This, therefore, carries with it connotations of censorship and punishment, as the
perceived silencing of users runs counter to values of free speech and open expression that
are so dear to online communities. Additionally, by positioning moderator action as an
exception to normality, any visible moderation becomes a sign of emergency or crisis. The
powers given to moderators are to be rarely exercised, and when they are, they carry with
them enormous anxieties over the proper use of power.

This is not to say concerns over abuse of power are not justified, but that the fear of this
happening is over-emphasized in the collective imagination of online communities. The re-
active model’s assumptions, that moderation work is punitive and exercises of moderating
power herald a crisis, means that all such action—when visible—is scrutinized. The codi-
fication of moderating rules or guidelines, by which moderators are meant to act, become
vital; objectivity becomes a virtue to which moderators should subscribe, mirroring popular
understandings of criminal justice proceedings. Through this prism, moderators occupy a
position of awful power. Unlike users, they possess administrator-like powers to remove
content and remove users, with the additional ability to read this other layer of invisible,
removed content; unlike administrators, they are not employed and therefore not clearly
answerable to the same hierarchies, and are much less distant and more visibly active within
the communities they govern. This uncomfortable liminality generates fear and anxiety, un-
derstandably, especially since moderation itself is meant to be invisible work. “Mod abuse”
becomes a rallying cry against threats of moderator overstep, real or imagined. Therefore,

23
containing moderators by holding them to standards of transparency and impartiality, gen-
erally accompanied by the aforementioned formal written rules of the community, becomes
imperative. Paradoxically, the need to have clear guidelines ahead of time is itself anxiety
inducing as this is akin to an admission that moderation will be required, which itself is an
admission of things going awry. Visible development of moderator policy, even without
accompanying action, is meant to come only when those policies are tested or in other such
emergencies.

The reactive model creates and sustains these contradictions, while at the same time
obscuring important elements of the relationships between users, moderators, and admin-
istrators. While the basic definition of users, moderators and administrators still holds, the
reactive model creates a strict boundary between users and moderators as a creative class of
users versus a reactive class of users. It also creates a binary between users and non-users,
since “moderator” in the reactive model is not generally clearly defined in the positive, and
more broadly means anyone with more permissions than regular users. Even within its nar-
row scope, it is ill-suited to explaining or accounting for moderator considerations directly
related to responding to content. For example, by tying all moderator work to direct re-
actions, short-term moderator action is disproportionately emphasized while longer-term
moderation work tends to be overlooked. Distinctions between different moderation roles
are also collapsed, as a consequence of the reactive model highlighting individual moderator
actions as the key area of focus.

Additionally, blurring the distinction between different types of moderator leads to its
own problems. Moderators are rarely recognized as a distinct group of users under the re-
active model. They are either lumped in with users, or assumed to be acting in the interests
of administrators. Thus, calls to expand volunteer moderation efforts are sometimes inter-
preted as calls to expand moderation powers to all users, as we can see with the classifica-
tion of visibility systems (such as likes or upvote/downvote systems), comment reporting
systems, or even recruiting users en masse as comment reviewers, as moderation systems.
Such efforts do not involve a distinct moderation class; in fact they actively blur the distinc-
tion between moderators and users. The vocabulary used for these systems may not reflect
the implementation across systems, either: on Facebook, reporting a post flags it for mod-

24
eration by Facebook’s own systems rather than volunteers; on Reddit, reporting a post flags
it for the volunteer moderators of that subcommunity.

On the other end of this extreme, moderation is sometimes held as the sole reserve of
administrators, who rely on opaque regulatory systems such as algorithmic content promo-
tion, automated content filters, and the like, often to cope with the sheer volume of content
that they must sift through. All of these regulatory systems are open to manipulation by bad-
faith actors. In the former case, their reliance on user input means they are vulnerable to
organized disruption. In the latter, case, their effectiveness relies on the precise calculation
by which these automated tools prohibit or promote content remaining hidden. As soon
as their mechanisms are understood, they, too, can be manipulated by organized groups to
promote specific forms of content, or to sidestep filters.

Because it forms the basis for our unspoken understanding of moderation work, the
reactive model has guided efforts to design and create moderation tools, or platform de-
sign more generally. For example, popular platforms such as Reddit, YouTube and Twitch
have by and large recognized the need for a distinct moderator class, to which community
members can be promoted by community leaders (the subreddit founder, channel owner,
or streamer, respectively). However, most platforms have been slow to adopt more sophisti-
cated moderation tools beyond the basic ability to remove comments or ban users. Facebook
Live, one of the more egregious examples, still has absolutely no affordance for moderators,
and it is impossible for anyone to ban users or remove comments on a Facebook Live video.
YouTube’s lack of granularity when it comes to assigning moderator permissions means that
channel owners face a difficult decision. Because editing and moderation permissions are set
for the entire channel, rather than, say, for individual contentious videos, or limited to cer-
tain actions, they must either give volunteer moderators almost total power over their own
channels, or choose not to take on volunteer moderators. This makes sense if the imagined
environment for moderation is a crisis situation, where it would be more expedient to give
an emergency moderator maximum permissions instead of having to hunt through more
granular settings. However, if moderation is part and parcel of a channel’s daily operation,
granting such wide-ranging power becomes more of a liability.

Using the reactive model as our basis for understanding moderation severely limits our

25
ability to collaborate with moderators (if this is even recognized as an option) and to create
tools to limit or combat online harassment. If we believe that all moderation is reaction,
then all the tools we create and the questions we pose revolve around faster reactions, rather
than seeing if there are other ways to pre-empt harassment and abuse. By focusing on the
event of removal itself as the be-all and end-all of moderation, we ignore the importance of
longer-term community care as we help repair the damage caused by harassment, and build
more robust methods for dealing with abuse of all kinds. We also ignore the fact that, by
positioning comment removal as the default solution to dealing with harassment, we also
default to letting the harassers get away with it, as if abuse were a force of nature rather than
a set of conscious choices made by other human beings. If we assume the correct course of
action is to remove an abusive comment after it has been made, we already accept that what
the harasser wants—for their abuse to be delivered to a public forum, or for it to be read by
their target—will always have already happened.
Lastly, the reactive model ignores constraints on moderator agency, as well as the way
in which nonhuman agents such as platform design influence moderator action. While it is
true that users lack the formal power that moderators possess, they are still capable of using
“soft”, social influence or other forms of resistance against moderator actions. Likewise,
moderators may take actions that run counter to the desires of some of their users, or counter
to the best interests of the platform’s administrators. Moderators may voluntarily constrain
themselves in accordance with ethical principles, or out of concern for the public fallout
of their actions. Even within the realm of reactive moderation, there is a complex web of
relationships, formal and informal, with their attendant tensions, considerations of power,
expected short- and long-term consequences, to factor in to every decision.

2.2 The lasting power of the reactive model

The popularity of the reactive model, as a way to conceive of the entire issue of moderation
online, is undeniable. To be sure, reactive content regulation work is a significant part of
what moderators do, and this has also shaped the ways in which moderators relate to, and
understand, their work. Content regulation is also the most visible aspect of moderation

26
work from both user and administrator perspectives. There are also practical considerations
that push moderators to hide themselves and the precise way their work functions from the
users they regulate. For example, certain content removal systems, especially those that
rely on relatively static, less-flexible nonhuman filters, need to be kept opaque in order to
be effective. Additionally, visible traces of moderation may hail crisis, but are themselves
a visible scar that disrupts the experience of other users. New technologies that hide these
traces (for example, removing even the “message deleted” notification) do provide a better
user experience in that it further minimizes the damage caused by the content that required
regulation. To put it another way, the crises and anxieties that visible moderation dredges
up with it are not necessarily unfounded.

The reactive model also simplifies the problems of moderation into a neater form, which
is more solvable. By portraying moderation work as a simple chain of cause-and-effect,
and focusing solely on those areas of moderation work that is the most conducive to this
portrayal—content regulation—the complex messiness of moderation work is cleaned up
and becomes a problem that can have a solution. When moderation work is no longer about
relatively fluid groups of people acting on, with, against and for one another, with multiple
motivations, abilities, valences and outcomes, it itself becomes something that is manage-
able, categorizable, and controllable. Pared down, it is easier to operationalize and automate.
A study conducted by Kou and Gui (2017), of players participating in Riot Games’ Tribunal
system, points out that for all its lauded successes, the Tribunal was quietly replaced by an
automated solution, despite the fact that “players repeatedly questioned the automated sys-
tem, citing its opaqueness, vulnerability, and inability to understand human behavior.” To
this day, the Tribunal system is still offline. Granted, we do not have information on the
efficiency of the Tribunal system later on, nor of its automated replacement, but the social
work that the Tribunal system performed—allowing participants to engage in moral labor,
whether or not this occurred because Riot “[convinced] people that it was righteous to par-
ticipate” (Kou and Gui, 2017)—does not seem so highly valued.

Acknowledging the social and affective dimensions of moderation, by contrast, means


acknowledging that human judgement is replaceable or reducible in the work, and that there
is a human toll on the workers who perform this labor. To frame moderation work as an

27
exercise of punitive power requiring objectivity and rationality is to place it within a hierar-
chical system, with users at the bottom, moderators in the middle, and platform operators
on top. I believe it would be a serious misstep to leave all oversight to corporations, or even
position them as the ultimate arbiters of behavioral regulation. Moderators and users, as
we shall see, have a robust understanding of and ways to deal with abuses of power even
in the absence of official tools or formal support for these deliberations. They are capable
of dealing with problems of power abuse and policy changes in a way that is attentive and
responsive to their needs, desires and values.

2.3 Counter-model: the proactive model

The proactive model grows primarily out of my experiences in community moderation on


Reddit. Though the particulars of moderating on different platforms are of course distinct,
the underlying ethos and domains of moderation work remain relatively consistent. The
proactive model is my attempt to expand the reactive model. The work of reacting to con-
tent does consist a significant portion of moderation work, and certainly moderators’ own
understandings of what they do is deeply influenced by this reactive work. Yet, this is not
the sum total of moderation work. While we may believe that this work is, or ought to be,
practically focused, objective, rational, and ultimately about the application of regulatory
functions by human or nonhuman actors, the day-to-day reality is more complex. Mod-
erators frequently engage in social and communicative work, coordinating between users,
fellow moderators, broadcasters or other personalities within their communities, and plat-
form administrators. They also engage in civic labor where they create and amend policies,
but also respond to policies or policy changes set forth by platform operators (Matias, 2016).
Their work is vital in creating a cohesive community (Silva et al., 2009) through the use of
soft social skills, not merely through the removal of un-permitted content.
This alternative model is less a replacement of the reactive model, and more an over-
haul of the same. Rather than think of moderation as comprised solely of discrete events
where moderators exercise regulatory power over the users they govern, the proactive model
places these exercises within a wider trajectory and backdrop of moderation. This trajectory

28
is formed from the interconnected practices, norms, behaviours, values, and affordances
found in the technical landscape in which both moderators and their community are sit-
uated, the social field of moderation (both specific to that platform, and broader cultural
considerations of moderation), and the reflexive social, mental and emotional work that
volunteer moderators conduct in order to comport themselves as moderators.

Furthermore, the temporality of moderation must be expanded to encompass the prepara-


tory work of moderators, and the ongoing social processes that make meaning out of reg-
ulatory work and fold it back into the service of changing or bolstering pre-existing social
attitudes and values regarding good or proper moderation. Moderators accumulate and
preserve knowledge gained from previous experiences, and are in constant dialogue with
their collective memory of these past exercises as it both guides and pressures current mod-
eration actions. Formal or otherwise, they remember and collect information on different
actors and use those memories in order to navigate moderation work. However, it is also
important to note that these are interpretive exercises: the relationship between past prece-
dent and current action is negotiated and interactive, mediated by the communicative and
archival affordances of whatever platforms and tools moderators can access, as well as the
values and norms of the moderator community in question.

Additionally, it recognizes that regulatory behaviour is not the only kind of work that
moderators perform. Under the proactive model, the technical work of developing, main-
taining, and adapting both in-built and third-party tools for moderation would qualify as
“moderation work”, as would emotional and mental health work conducted by moderators
for their communities and for each other. Lastly, it complicates the position of moderators
within their existing networks by acknowledging the impact of other groups, such as users,
external organizations or corporations, platform operators, and other relevant parties, on
moderators and their behaviour.

This more holistic understanding of the work of volunteer moderators uncovers their
invisible work, in order to better appreciate the work already performed by these actors.
Without a broader understanding of volunteer moderation work, efforts to improve social
spaces online may well fall short, as we neglect a key group that already has a robust history
of employing, creating and adapting whatever resources are available to them in order to

29
perform this kind of community-building work. With a better knowledge of what it is that
moderators do, we can better create the infrastructure and resources necessary to support
them in performing this crucial labor. The invisibility of moderation work need not remain
unacknowledged, even as it remains largely unseen.

30
Chapter 3

Moderation on Twitch

Twitch.tv is an immensely popular livestreaming site, which allows its users to host live video
of themselves. Launched in 2011, Twitch has its roots in Justin.tv’s games livestreaming sec-
tion, before it became big enough to split off. Twitch is easily the dominant livestreaming
platform for esports and other major game tournaments, although recent efforts by other
companies such as Google (through its YouTube Gaming program) and Facebook have
emerged as alternatives.
Twitch is split up into different channels, each controlled by a single streaming account.
Viewers largely interact with streamers through a live chat panel that can be found to the
right of every single stream. As viewers on Twitch settle on particular channels that they
deem their favourites, communities form. These same communities may come together to
view one-off events, exchanging norms, ideas, and on the most superficial level, different
memes, emotes, and new ways to spam messages in chat.
Twitch chat is the site of complex interaction between streamer, chat user, and moder-
ator. Through repetition and iteration of simple, easily copied-and-pasted messages (com-
monly called “spam” in the parlance of Twitch), the effect of a crowd of fans roaring for their
favourite teams is replicated. However, spam is rarely so straightforward: in-jokes abound,
and remixes of existing spam to change the meaning are common. Moderators on Twitch
ostensibly are primarily charged with managing this often unruly crowd, regulating the types
1
of speech found in chat.
1
to be expanded

31
Figure 3-1: A screenshot of Twitch.

The two most commonly seen types of volunteer moderation on Twitch are community
and event moderation. It is important to note that these forms of volunteer moderation are
not diametrically opposed, and oftentimes the same person will take on both roles on an
as-needed basis.

Community moderators are volunteer moderators that typically take care of a single
streamer, or network of like-minded streamers. In addition to regulatory work, they may
take on additional community management work. For example, a community moderator
may take it upon themselves to greet newcomers to the channel, or to redirect viewers to
different resources to learn more about the stream. This might expand out to managing
auxiliary fansites, or cultivating different specializations: a moderator might be responsi-
ble for producing custom graphics, or managing different bots, tools, or scripts used in the
channel. Because the purpose of establishing a channel on Twitch is to cultivate a stable
repeat audience, and to steadily grow it, moderators end up forming relationships with the
viewers, especially channel regulars. The flow of chat is also more of a back-and-forth be-
tween the streamer and their viewers, producing a more conversational atmosphere. The
lengthier broadcast times might also lead to more hours worked per moderator, on a more
consistent basis than event moderators.

32
As previously mentioned, event moderation is centered around discrete events that typ-
ically last no longer than a single weekend. Chat during event livestreams are rarely conver-
sational, mostly consisting of audience reactions to things shown on-screen, or repeated fan-
chants, memes and spam, due to the sheer volume of comments. Events draw high viewer
counts, but these viewers are unlikely to stick around and form any basis for a permanent
community. Instead, they coalesce for each event. While there is some overlap between
different community streams and events (for example, the regular viewers of a esports pro-
fessional may be likely to chat during events where that professional is participating), events
are rarely dominated by a single community.

3.1 Moderation tools

In order to expedite these kinds of mass moderation work, moderators turn to different
tools, many of which are developed by third parties and far exceed what the platform it-
self offers in terms of moderation functionality. To generalize, moderators rely on tools
modify Twitch’s user interface, add in chat history or archival search functions, and create
flexible automated chat filters that can be adapted on-the-fly as new situations arise. Mod-
erators also use other applications and programs, for communication, chat monitoring, and
other peripheral considerations. They also employ built-in tools and settings available in
Twitch itself. Third-party tools do not wholly supplant Twitch’s built-in tools, but greatly
expand what they are capable of accomplishing on the platform, and some were considered
by my interviewees to be indispensable to their work. These tools, however, are not officially
supported by Twitch; they exist in a kind of careful dance around the official updates and
rhythms of the platform, and every update is scrutinized by the maintainers of these tools
as they represent new points of failure for their tools. This is the case even for updates that
promise to help moderators by building on Twitch’s inbuilt moderation options.
As Twitch’s moderation tools developed, so have these third-party tools correspondingly
grown more complex and powerful. Similarly, tools that are widely adopted by the modera-
tion community and championed by respected individuals gain prominence and popularity.
These tools should be understood as created and informed by the practices of moderators

33
paired with the relatively new ability for individuals or small teams of developers to use plat-
form APIs, such as Twitch, to make tools that tap into those systems directly. Twitch’s open
(or more open) API afforded the growth of a moderation tool ecosystem, which in turn
supported the growth and sustenance of a class of moderators. While an open API is not a
necessary condition for a healthy moderation community to form, it is an important factor.

On extremely powerful tool moderators have at their disposal are bots. These are pro-
grams that automate many of the more common actions that moderators would be expected
to perform; on Twitch, there exist many different bots aimed at different groups of users,
with the majority aimed at helping out streamers. Bots that prioritize the needs of modera-
tors are rarer, though many streamer-facing bots have features that make them suitable for
moderation as well. Common functions that these bots perform include removing messages
and users, permanently or temporarily, according to certain criteria, or dispensing informa-
tion through the use of custom chat commands. Bots are often named after their creators,
and are funded through many different models: some are wholly free, while others might
have premium features locked behind a one-time or recurring payments. Of the moderators
I interviewed, the two bots most commonly named were moobot and ohbot.

Moobot has a well-maintained graphical interface, allowing moderators to change its


settings via an external dashboard. It can be programmed with custom commands, for ex-
ample allowing users to message moobot in order to get schedule information for the event
they are watching. Moobot also has an automated spam filter that stops messages with excess
capitalization, punctuation, repetition, certain memes, as well as allowing a custom blacklist
of phrases or words. It is donation-funded; while no donation is necessary to use its basic
features, donating money to moobot gives a user points, with which they can unlock more
features, and more slots for editors. Editors are users who have access to moobot’s settings;
for a moderation team, this might include anyone trusted with changing blacklist phrases
or adjusting its filters.

Ohbot, by contrast, is far more difficult to set up. It has no user interface, meaning
that moderators can only change its settings by typing commands into chat. However, it is
one of the very rare bots that is meant primarily for moderation. Its primary function is as
a chat filter, and it not only allows for extreme granularity in settings but also can match

34
strings using regular expressions, or regex. Regular expressions are search patterns which
allow for far more powerful search and pattern-matching capabilities than normal word or
phrase blacklists. Using a standard syntax, regular expression strings can catch many vari-
ations on the same word or phrase, which means that a single well-tested regex string can
have the same effect as multiple blacklist entries. Regular expressions can also be used to
check the context in which a phrase is used, since it can check ahead or behind the phrase
in question. For example, a regex string could be set up to permit ‘TriHard’, a global Twich
chat emote depicting a black man. However, that same string could be set to match if ‘Tri-
Hard’ was embedded within a longer, racist message. Equally, a mispelled, poorly-tested
or poorly-thought out regular expression string also has the potential to cause trouble. A
misconfigured regex string might end up matching not enough or no messages, rendering
it useless as a chat filter. Or, it could match too much, and incorrectly ban or timeout users
sending innocuous messages, forcing moderators to reverse these bans, mollify chat, and
take down the regex filter to be fixed.

Mastery of regular expressions is a specialist skill that is not common in all moderating
circles, and those who understand regex are sought out by head moderators. Moderators
who are responsible for creating these regex strings are guard them closely; one head mod-
erator said that their regex string was seen by “about 6 people”, all of whom had access only
because they were also active contributors. One particularly well-known moderator’s regex
settings are now part of ohbot’s presets, and the presence of a name helps prove its efficacy
by explicitly giving that preset a respected author. However, not all moderators seek to learn
it, simply because it is quite complicated, takes time and effort to learn, and comes with a
different set of responsibilities. In the words of one moderator, “I’m good at what I do now,
I don’t want to to much more than that. I don’t want to pick up everything, to do all the
botwork and the programming behind that. I’ll pass on that one.”

Many other bots exist, with slightly different sets of features. The moderators I inter-
viewed seemed to choose which bots they used based on their own familiarity with them,
their moderation team’s familiarity with them, the features available to their chosen bot, and
whether or not the bot was in active development. Some of the other bots mentioned by my
interviewees were xanbot, hnlbot, or Nightbot. One moderator I talked to also worked to

35
develop their own bot, to allow them to check accounts by age.

There is another bot available to all moderators, AutoMod, which is built into Twitch.
AutoMod holds messages for human review, and uses machine learning to determine which
messages should be held. AutoMod has four settings, which increase its sensitivity and
change the types of messages that it targets; on the lowest, least-sensitive setting, it might
only filter out racist slurs, while on the strictest it will remove all forms of hate speech, vi-
olent language, and any profanity. Though the moderators I spoke to appreciated it, espe-
cially once its had matured a little past its debut performance, they did not regard it as a
one-size-fits-all solution to chat filtering. AutoMod can catch most general use forms of
impermissible speech, but is relatively easy to circumvent by using emotes, memes, racist
stereotypes or scene-specific in-jokes to express the same offensive sentiments. Modera-
tors cannot respond by setting AutoMod to a higher setting, for two reasons; firstly, only the
broadcasting channel account can change AutoMod settings; secondly, since the settings are
relatively opaque and come in bundles, setting AutoMod to be more restrictive risks chilling
chat to a degree deemed unacceptable by the moderators I spoke to. In my own observa-
tions, though AutoMod did filter out many messages, there were a significant portion of
messages that were removed either by moobot, ohbot, or direct human intervention.

The second most common tool mentioned by my interviewees was Logviewer. This is a
quasi-bot tool, which sits in channels that have opted in and generates a log of all messages
that have been said in it. Moderators then log into an external site with their Twitch accounts,
granting them access to the full chat history of the channel. Crucially, Logviewer allows
moderators to see an individual user’s chat history within that channel for as long as it has
opted into Logviewer. Moderators can also add comments on a user, which can be viewed
by all the other moderators.

According to my interviewees, Logviewer is useful because it allows moderators to keep


a record of events. This is most useful when handling unban requests, or trying to sort out
disputes between users. Since it also records the name of the moderator who performed
bans, timeouts or other actions, the team ccan also determine who should be a part of any
decisions for unbanning users. It is so useful and popular that it has been integrated with
another popular third-party tool, FrankerFaceZ, and there exist other, even more special-

36
Figure 3-2: An example of Logviewer, showing multiple user chat histories, with moderator
comments on a user.

ized bots that link Logviewer to other applications such as Discord. A similar logging tool,
Modlog, exists specifically to track moderator actions. It allows head moderators to see who
has been working and when, as well as to spot suspicious activity that might indicate a com-
promised account. Modlog similarly has bots that link it to Discord, so that moderators who
are not present in Twitch chat are still notified of important moderator actions that a given
team might want to review.

The last tool mentioned by all my interviewees was some sort of user interface improve-
ment, and the two that were named were Better Twitch TV and FrankerFaceZ, or FFZ. Both
are browser plugins, and essentially give users the ability to customize the way Twitch chat
looks. Though neither was developed primarily for moderators, they are nonetheless widely
used for this purpose. Of the two, FrankerFaceZ was preferred by the moderators I inter-
viewed because of its more active developer, and because it performed all of Better Twitch
TV’s functions as well. FFZ’s developer has also collaborated with the maker of Logviewer,
to integrate Logviewer into FFZ, making it even more helpful for moderators.

These extensions primarily make it easier for moderators to perform common actions,
such as timeouts and bans, and to view more information in one space. Figure 3-3 shows a

37
Figure 3-3: FrankerFaceZ’s moderation card.

‘moderation card’, which FFZ brings up if a moderator clicks on a user’s name. The moder-
ation ‘card’ can display the user’s chat history, and allows the moderator to choose from a
series of timeouts (from five minutes to one week) in addition to adding a drop-down menu
with a list of customizable ban reasons. It also adds hotkeys for timeouts, bans and purges,
and can highlight messages that contain a particular phrase in chat. This, in conjunction
with pause-on-mouse-hover, allows moderators to perform moderation actions at a much
faster rate, and to keep up with the pace of Twitch chat..
Because chat moves by so fast, FFZ’s user interface changes shine for moderators in large-
scale chats. One of the most beloved features it adds is the ability to pause chat on mouseover,
with one moderator calling it “a godsend,” adding, ”we also mod on YouTube, and I literally
didn’t mod YouTube for the past year and a half because you couldn’t slow down chat, and
things would just fly by and unless I could scroll up as fast as the chat was going it was just
impossible.” Issuing a timeout or ban involves either typing in the appropriate command
in chat—/timeout username [duration] or /ban username—or clicking one of two
small icons by a user’s name. Without the ability to pause on chat, misclicks are very com-
mon and necessitate fixing, causing more work for moderators.
Lastly, in order to communicate with each other, moderators set up moderator-only
groups where they can discuss moderation policy with each other. Since Twitch is set up to
revolve around a video stream with accompanying live chat, it is not ideal for private group
community discussions about individual moderation decisions. Instead, common ancillary

38
platformed used by the moderators I interviewed were Slack, Discord, and Skype. Skype
has fallen out of favor with the advent of Discord and Slack, with moderators citing secu-
rity concerns as a reason why they tend to shun organizing on Skype. Discord and Slack
also have support for markdown, image and file sharing, and multiple text channels. These
features allow for rich text formatting and easy sharing of screenshots or other files which
make communication easier. Discord seems more popular for its gamer-oriented branding,
granularity of permissions (through a role system, which also makes it a little easier to set
up a moderator hierarchy), ability to set up voice chat, and support for bots using its API.

Within this third-party tool ecosystem, there are other artifacts that are not designed
to directly help moderators work. Over my time observing this community, I saw a slew
of other peripheral technical solutions, created to prop up this ecosystem. These included
fixes, workarounds, add-ons, and other such kludgey solutions. These are not necessarily
solutions meant to last, as the development cycle of both Twitch itself and these third-party
tools rapidly forces them into obsolescence. However, instead of dismissing them as tempo-
rary elements meant to fade away over time, it is more useful to understand them as doing
bridging work (Braun, 2013). These systems should also be seen and understood as part of
the moderating tool ecosystem: involving the same actors, along the same networks, with
impacts that may outlast the time in which they are in common use. The users who create
tools for long-term maintenance may well be the same as those who push out quick fixes to
make them work with the latest versions of Twitch, and add-on functionality may some day
be incorporated into the tools that they enhance. Some examples of these bridging appli-
cations include tools that link Discord with Twitch, so that moderators can be working on
Twitch, or monitoring it, without ever actually opening the site itself, and a Twitch Legacy
Chat application, rolling Twitch’s chat back to an older version supported by popular tools.

Aside from these plugins, applications, bots and scripts, there are a smattering of oth-
ers that do not have a clear place within this complex constellation of tools. For example,
Twitch requires the use of a third-party application, Authy, in order to set up two-factor
authentication on one’s account. Taking such security measures was something all the in-
terviewed mods did, and so they all had to use Authy. Other useful moderating tools include
Multi-Twitch, which allows a user to display multiple chat windows as well as multiple video

39
streams side-by-side in a single window.2 Custom chat clients, such as 3ventic’s Chat Client
(designed for moderation) and Chatty, are also sometimes used, though they are not quite
as flexible as FFZ. Moderators may also have mass unmodding or mass unbanning scripts
on hand in case of compromised moderator accounts. However, my interviewes indicated
that with the introduction of better security features such as two-factor authentication for
accounts, the need for these drastic measure has lessened.

The value of these tools is most evident when they are absent. During IEM Katowice, for
which I was a junior moderator, the moderation team decided to try and use the new Rooms
feature of Twitch. Introduced in February 2018, rooms are side-channels, and allow a stream
to subdivide its chat between a main chat and smaller peripheral rooms.3 For this event, two
rooms were set up, one for each team playing. However, this update had also broken nearly
all of the moderation tools I described above. Only FrankerFaceZ had limited functionality,
and even then, none of the moderation features were working. To get around this, the mod-
erators installed another plugin which forced Twitch to use ‘legacy chat’, an older version of
the chatrooms, which would be compatible with their tool suites. This had the side effect of
making the new rooms invisible to both moderators and most of the moderator bots. Over
the last day of the tournament, I spent most of my time sans tools, working only with the
built-in moderator functions that Twitch provided.

Although the rooms were slower-paced than the main channel, it was still difficult to
watch and regulate. If a single user was responsible for spamming the channel and mak-
ing it impossible for others to participate, I then had to type in the appropriate timeout
command as quickly as possible, with an appropriate duration in seconds. At a very basic
level, this required parsing chat at least as quickly as it flowed, very fast and accurate typ-
ing, and memorizing different minute durations in seconds and making a judgement call to
match the behaviour to a timeout length. If multiple users were participating in unpermit-
ted behavior—for example, if one user encouraged many others to raid the opposing team’s
room by spamming insults to make it unusable—I had to call in help on the moderator-only

2
See figure 3-4b.
3
Interestingly, in the blog post announcing this new feature, Twitch suggested that moderators could use
it as a private communication space. However, none of the moderator teams I spoke to had even considered
doing so.

40
Discord room. Without the ability to look up user chat histories in the rooms, the modera-
tors had to rely purely on their memory of events when adjudicating unbans after the event,
generally erring on the side of caution.
Understandably, the presence of these tools was greatly appreciated by the moderators I
interviewed. The ubiquity of these tools, and the streamlining they provide for moderation
work, is so drastic that I would say that there is a generational gap between newer moderators
who are used to these automated tools, and moderators who learned how to moderate in the
absence of these tools. This sudden reliance on scripted tools, and the subsequent valuing
of these new tools, did not seem to lead to jealousy or resentment in the moderator circles
I studied. Instead, the prevailing atmosphere seemed to be one of gratitude, recognizing
how much easier these tools had made online moderation. One self-described “old fart”
moderator described this as “They [newer moderators] incorporate what they are studying,
for example, into moderation…I see these young guys and I try to include them within the
community more, and suddenly they rise way above me, for example in title, in certain areas.
That’s just amazing to watch, because not only did they make the entire community better,
but they’re getting the respect that they deserve.”
It is important to remember that the explosion of moderation tools for Twitch did not
arise ex nihilo. In the words of one moderator, “what the platform lacks is always that which
the community creates itself.” A common theme was that all the tools available to modera-
tors were developed out of necessity, to fill gaps in Twitch’s built-in moderation tools. Early
on, this was done in the absence of these tools; however, the creation of more sophisticated
functions such as Twitch’s AutoMod has not diminished their importance in any way. In
fact, one moderator stated that they were working on a bot that would have performed a
similar function.

I was teamed up with a friend to make a bot. The point of the bot was to analyze
messages in channels, and determine what sorts of messages result in a ban, us-
ing machine learning [to] begin estimating what messages we should ban, and
if that becomes solid enough, release that model and start banning messages us-
ing the bot, or warning mods that those messages should be banned. Literally
a month after we started working on this project, Twitch released AutoMod,

41
which essentially does the same thing.

The development of these moderation tools is made possible by the affordances of the
platform, the ease by which moderators can learn to create them, and the practical experi-
ence and knowledge generated by moderators guiding the creation of moderation tools. It is
clear that moderators are not bound by what moderation options are built into the platform;
rather, they are more restricted by how open said platform is. A few of my interviewees said
that they disliked moderating on Youtube’s livestreams, partly because of moderation input
lag, but also because the moderation ecosystem was tiny in comparison to Twitch’s.4
I interviewed two moderators who created and maintained third-party tools, and both
of them mentioned that they had created these tools for themselves, and realized the need
for these functions once large numbers of moderators asked them for access. When asked
if crowdfunding platforms like Patreon might affect the mod-tool landscape in the future,
both of them were skeptical, pointing out that even the most successful moderation tools did
not bring in any appreciable profit for their creators, and even the most successful moderator
Patreons would just about cover server and operating costs. With little monetary motivation
to create and sustain these tools, it seems that the primary driver behind moderation tool
development is the need and desire for better moderation tools, created by moderators for
this relatively small group.

3.2 Running an event

Moderators do not simply show up at the start of an event and begin to remove messages; in
an ideal situation, they will have prepared for this days in advance. Each event should also be
understood as an opportunity for new moderators to learn how to moderate, for more expe-
rienced moderators to refine their skills and style, and for those moderators implementing
new tools to test out their creations. Each event also has its own specific requirements and
regulations coming from event organizers and the makeup of the presumed (and actual) au-
4
Though Youtube’s live streaming API has been available for many years, the moderators I spoke to only
mentioned one moderation bot of note, Nightbot. Nightbot itself started off as one of the first Twitch moder-
ation bots.

42
dience, necessitating on-the-fly adjustments by moderators. What follows is a generalized
and ideal account of the steps involved in moderating a large-scale esports event on Twitch.

3.2.1 Preparation

In the run-up to an event, the head moderator may be contacted by a company liaison with
relevant information. The head moderators I talked to said that this liaison usually was the
social media manager, but ideally would be anyone with access to the broadcasting account
on Twitch and contact with the production team. Head moderators pass on the account
names of those who have volunteered to act as moderators, since the broadcasting account
must be the one to make them moderators for the channel. The company liaison, in turn,
gives head moderators information about the tournament schedule, predicted match start
times, the teams in the brackets, the names and Twitter accounts of the casters, commen-
tators and analysts, and other miscellaneous information such as the names of the songs
that will be played during the event, or any promotional events such as giveaways that will
happen during the tournament.
The moderators then work to set up the bots for the event; for the moderators I inter-
viewed, this consisted of setting up moobot and ohbot. Moobot is a donation-funded bot
that has a graphical user interface, making it much easier to use. It is capable of setting fil-
ters on chat, such as punctuation filters which stop messages that have over a certain ratio
of punctuation-to-message-length, capitalization filters, and an easily-changed blacklist for
phrases or words. It can also be programmed with custom chat commands, which can be
triggered either by moderators or users. The head moderators, or any moderator with ac-
cess to moobot ahead of the event, will wipe previous blacklists or custom commands so as
not to trigger false positives when the event starts. New custom commands for displaying
the schedule, score for the match being streamed, song titles, analyst information and more
will be added using the information received from the company liaison. These generally
take the form of !command, which users can use in chat to get that information. For exam-
ple, a viewer who wants to know the day’s schedule could type !schedule in order to get a
message from Moobot that would link them to a webpage with the day’s schedule.
At the same time, head moderators look for people to help work the event. Moderators

43
might be approached individually on recommendation, or recruitment calls might be posted
in moderator community spaces. The recruitment process is fairly informal; the measure of
a new moderator seems to be in the amount of work they do and their willingness to remain
an active moderator, with a moderator’s reputation in the community serving as a major
factor in whether or not they find more moderation work. Newly recruited moderators are
brought into event-specific moderator chat groups, and are told what this particular event’s
guidelines for chat are. Not every organization or developer will be the ones to issue such
guidelines, and one head moderator said that “most people just leave it up to us [modera-
tors], that have been doing it the longest.”
There is no set consensus on how large a moderation team should be for an event of any
given size. In fact, some moderators say that a few—numbers range from two to five—very
active moderators are sufficient even for events of over 100,000 viewers. “The ELEAGUE
[CS:GO tournament] was the biggest thing ever, a million people watching. And it was prob-
ably four of us active. There were a lot more people there, but only four of us were active.”
Other moderators concurred, and the phenomenon of a few active moderators managing
an entire chat kept resurfacing. “there are two or three moderators who do 70% of the stuff.
It’s always like this. Most of the time it’s the same people, the same very active people…and
maybe five more who are semi-active, and maybe ten or fifteen others who do 1% each.”
However, head moderators do try and recruit moderators with different availabilities
in order to cover as much of the event as possible. Head moderators often try to recruit
people from different timezones, which can be difficult for certain timezones when looking
for English-speaking moderators.

It’s a typical thing, that it’s very hard to find people from the Oceania or East-
ern Asia regions…There’s so few people that know English well enough to do
moderation, there’s so few people from that region in the community itself, it’s
basically just people from Australia. The population is so low that—it is ridicu-
lously hard to find people from there!

Equally, a moderator might prepare themselves for the event by clearing their personal
schedules. One of the more extreme anecdotes I encountered involved a moderator who

44
would “just change [their] sleep schedule” to better sync up with the event schedule. Another
offhandedly mentioned that “since I have insomnia, I’m there for the European stuff no-one
else can do.” In any case, moderators coordinate to tell each other when they are available
(or not) to monitor chat, with the intent to have some overlap to allow moderators to leave
to take care of themselves, or simply take a break to focus on the games rather than watching
chat alone. These events are huge time sinks, and over the course of the weekend a moderator
may well spend upwards of ten hours per day on the event. Part of this is that moderators
need to show up early to the event, because some viewers will show up before the stream
starts, and viewers will stay after the stream ends, which means some moderators must keep
an eye on chat after its official stop as well.

It’s more than a full time job during an event. I track my uptime, the hours I
spent working on [a large recurring event] for two years, 2016 and ’15 I think?
I usually had roughly 140 hours of uptime in an event that ran for a week, for
seven days…I [needed to] sleep, and I had university, I was up for all but like
twelve hours. I slept for four hours a day at best. It was stressful. It’s why I
always suggest having more than one head moderator, because a single person
can simply not take the load. Especially for a 24/7 event, it’s just impossible.
And if they don’t take the load, or if there’s not enough people, then they burn
out and the quality of the management goes down.

3.2.2 During the Event

When the event proper starts, moderators load up different tools and applications. Com-
mon ones include Discord, opened to the event moderator group; a window with either
the Twitch page itself, or one of several moderator-focused Twitch chat clients that enable
multiple chats to be placed side-by-side in the same window; control panels for bots like
moobot; Twitter or Reddit to monitor other chatter about the event; and personal programs
such as a music player or a simple text editor.5 This enables moderators to respond quickly
to queries, to monitor multiple simultaneous chats—useful if the same video has multiple
5
See Figure 3-4 for some sample screenshots of a moderator’s setup.

45
alternate streams, as is the case with some major tournaments—in one window, to change
bot settings on the fly, and a text editor allows a moderator to catch new spam as it flies by.

What mods are looking for are patterns within the chat. These patterns generally come
as memes or spam. Spam was not well-defined by the moderators I interviewed, but can be
taken to mean any message intended for multiple users to copy and paste to as to quickly
repeat it in a very short span of time, while memes are a broader category of short, repeatable
message. Spam in esports chat often involves the use of global or FFZ emotes, juxtaposing
both the surface level meaning of the emotes-as-images with what they mean to esports
fans, with some accompanying message, whether done in regular text, ASCII, or Unicode.
Twitch esports chat spam changes extremely rapidly. Some spam is a response to the action
happening on stream; others might be akin to fan chants, proclaiming support for one team
or another.

Moderators are not paying attention to what individuals are saying in chat. Rather, they
are paying attention to the patterns and rhythms of chat: which messages get repeated, if
they are relevant to the stream or if they are irrelevant spam, and most crucially, if they
are offensive, disruptive, or otherwise exhibiting a moderation “red flag”. Over time, ex-
perienced moderators are able to predict which spam is most likely to get picked up and
repeated. “When you watch chat for so long you pretty much know what’s going to hap-
pen when it happens. It’s like the Twitch chat whisperer, you just know everything.” This
sense for Twitch chat allows more experienced moderators to warn newer moderators what
to look out for, and for them to react quickly to head off spam that might encourage more
and more-offensive messages from being posted.

The types of offensive spam that moderators encounter are ever-changing. The most
easily-identified offensive spam feature slurs or other offensive terms. These terms might
be spelled plainly, or by using letter-substitution via using similar-looking glyphs, Unicode
or ASCII symbols. More involved offensive messages involved the use of ASCII or Unicode
to create offensive images, or the use of emotes (default, subscriber-only, or added by ex-
tensions such as BTTV or FFZ) to create offensive phrases and imagery. Moderators also
described a class of offensive spam designed to trick viewers unfamiliar with American cul-
tural norms into repeating offensive phrases, showcasing the sometimes creatively devious

46
nature of this spam.

Twitch chat has a thing where it tries to trick other people into being racist in-
stead of the original commentor. It’s basically trolling people into being racist…As
with all groups of people if it becomes too large, it’s just a mess of idiots, and
Twitch chat, if it sees something, it will most likely copy-paste it within sec-
onds…There’s been times where I’ve banned people for saying horrible things,
and the person actually didn’t know what they were copy-pasting at all.

For popular spam written in languages other than English, a moderator may try to tap
others with other language skills in order to ensure spam remains slur-free. In my obser-
vations, I saw calls for Russian-speaking moderators, as well as German, Czech and Hun-
garian; knowledge of multiple languages remains a valued skill for moderators, especially as
this sometimes corresponds to wider timezone coverage. Otherwise, moderators use tools
such as Google Translate to try and get a quick sense of what the message means. Some non-
English spam is so common that moderators who do not otherwise speak those languages
do not need to check its translation to know what it means, and know when to remove it.
A large chat may be set up with several modes aimed at fractionally slowing down chat
and protecting it from the most easily exploitable methods used by spammers to dominate.
This would be turning on slow mode and a chat delay for moderation. Slow mode means that
each user can only post messages after a certain time delay, preventing a single user or group
of users from quickly overwhelming chat. Chat delay means that all messages have a small
time delay, with moderators seeing them before they are passed to the public stream. This
small delay allows moderators, and bots with moderator status, to act on messages ideally
before they are even seen.
The first barrier to disallowed chat messages is AutoMod, which catches messages in
its filter and prevents them from hitting chat at all, instead passing them to moderators for
manual approval. Notably, a message held for human review by AutoMod will not show
up for users at a whole, meaning that even without anyone giving AutoMod explicit ‘allow’
or ‘deny’ feedback, there is some level of chat filtering that will occur. The second is for a
moderator, on spotting a disruptive message, to manually add it to a bot blacklist. This first

47
short-term action is generally done in a user-friendly bot with a GUI, such as moobot. There
are two reasons for this: a GUI may allow for faster response as it can be opened and readied
in a separate window, some mod hierarchies are organized such that most mods will have
edit permissions for these less-powerful but easier-to-operate bots, and halting the exact
spam message quickly limits its spread.6 The longer-term solution, if the blacklist does not
halt the spam, is to work this message into more powerful filters using regular expressions.
Badly-configured regex strings have the potential to filter too many messages, turning the
chat against the moderators, or to filter too few, allowing prohibited messages through; such
was the case with one observed incident, where a typo meant that a regex filter meant to catch
messages using a certain combination of punctuation was, in fact, catching no messages at
all.
If possible, other moderators then encourage chat to continue by engaging in non-offensive
memes or playing games with chat. This may also be done during down-time, such as sched-
uled ad breaks, or technical issues. This can also be a planned and delegated role; if modera-
tors know a minority figure will be on screen, they may set up specific filters (for example, if
a female commentator is part of the broadcast team, they may prepare stricter filters target-
ing sexist language) and be prepared to deploy a moderator specifically to distract chat when
they are on-screen. This distraction generally involves trying to steer chat in the direction of
less spotlighting and more benign behaviour; one example given to me was trying to engage
chat in the ‘golden Kappa test’ or to ask them to ‘build a rainbow ladder’. Both of these are
well-known Twitch chat games, which essentially demonstrates participants’ knowledge of
Twitch by trying to trigger a known easter egg or by using special ASCII characters, respec-
tively. The intent is to divert attention from the screen and to give chat something to do
other than point out the minority status of the person on screen. Simultaneously, the users
that originated this now-blocked spam may be timed out or banned, depending on severity.
Once the new filters are in place, moderators look out for signs that users (particularly
those that originated the spam) are testing the boundaries of the new filters. One moder-
ator, who said they were normally loath to “make an example” of a user, singled out these
6
Few of these methods are backed up by data or evidence proving their efficacy, but through practice and
observing the outcomes of adopting certain moderation actions over others, Twitch moderators have devel-
oped sophisticated techniques to regulate chat.

48
boundary-testing users as particularly dangerous “since they know how we blacklist things,
they know how to edit it just right so it’ll get through again. Those are the people I throw
a day ban on. I’ll @ them7 after I ban them, ‘Congratulations, you found a way to get past
the filter, here’s your prize.’ That way chat knows they shouldn’t do it.” Moderators are very
harsh towards these kinds of probing attempts, since they signal a conscious and deliberate
attempt to test the abilities or attention of the moderators. If the moderators are sure that
the popularity of that particular piece of spam has waned, it may be removed from all black-
lists to stop false positives. Again, it is important to keep in mind that it is not necessarily
the content of the spammed message that warrants its removal, but the fact that it is being
repeated so much that no other messages can get through.
In a particularly dire scenario, moderators may limit the rate at which messages can be
posted, using a more restrictive slow mode, r9k mode, a stricter AutoMod setting, follow-
only mode, or sub-only mode. r9k mode disallows repeated identical messages, which
forces spammers to change their messages before posting and thus slowing the flow of spam.
Stricter AutoMod settings catch more and more messages, although an overly-strict Auto-
Mod setting can actually become detrimental since it is not nearly as configurable as either
ohbot or moobot; higher settings both catch more messages and broaden the net to catch
more types of content. Follow-only mode restricts chat to those who have followed the chan-
nel for some amount of time, which slows the flow of newcomers jumping in and also stops
banned users or other bad-faith actors from creating new, throwaway accounts to spam the
channel. Sub-only mode only allows subscribers of the channel, those who pay money to
support the stream, to chat. This is seen as equivalent to turning off chat altogether, since
so few users subscribe to event channels. Because chat is the primary source of interaction
between users and an event, and is vital to the cultivation of an organization’s brand and
fanbase, turning it off is a last resort.

3.2.3 Cleanup

Once an event ends, some moderators might stay around until the majority of the users
have actually left, since it is common for people to hang around after the stream ends. Aside
7
On Twitch, typing @username pings a user with that username, highlighting the message for them.

49
from that, there is little to do after the event. “The head mod will always be like, ‘Good event
guys, thanks a lot, our next event is going to be at so-and-so.’…At the end of the event, we go
through all the commands in moobot and disable them so the people don’t call in the chat,
they don’t prompt the command and get information from a month ago.” However, the bulk
of post-event work comes in the form of dealing with unban requests. Moderators may get
private messages on Twitch from users that have been banned; since users are not always
told who has banned them, the chances of a banned user contacting the moderator who
was responsible for their ban are extremely slim. With a tool like Logviewer, any moderator
can check a user’s history and see what messages they posted up to their ban, and see if
moderators have left additional notes about that user. Without access to Logviewer—for
example, when I participated in moderation, a recent Twitch update had broken Logviewer
functionality for the new Rooms feature—moderators must rely on each other’s memories of
the event to decide whether or not to unban a user. With luck, a moderator might have had
the foresight to keep logs, generated from some other chat client, of the event; post-event
logs are useless as deleted messages are not preserved even for moderators.

Moderators may also experience an uptick of angry messages left by irate users. Learn-
ing to handle these among moderators is a little-discussed part of becoming a moderator, as
the practice of sharing these angry and sometimes threatening messages is the most com-
mon form of emotional and mental care in moderator circles. Some users are persistent:
“we laugh about the death threats. There was, somebody posted death threats to almost ev-
ery single ESL mod at an event two or three weeks ago and we were laughing about it in
mod chat.” Though moderators know they are not meant to share private messages, in these
circumstances it is very common and forms the basis of moderator-specific memes and hu-
mor. Counter-trolling or baiting these angry users is considered an acceptable response.
“Sometimes it can even be funny because they’re so aggressive for things that are so negli-
gibly unimportant. Often what I did was I had my fun with them, I pretended to be super
nice.”

Compensation may be given to the moderators. This is so rare as to almost be unheard


of, and even then almost never comes in the form of money. If compensation is given,
moderators may be given tickets for future events with VIP access, invitations to exclusive

50
afterparties, or “swag bags” of branded merchandise. Some moderators, who self-identify as
fans of the esport for which they work, view the opportunity to meet esports celebrities as a
form of compensation in itself. However, even these non-monetary forms of compensation
are still incredibly rare.

3.3 The reality of event moderation

The section above assumes an ideal scenario, but not every event goes as planned. To begin
with, not every event that is streamed on Twitch has a human moderation team, let alone a
moderator liaison. With the development of AutoMod and the wide availability of automate
moderation bots, such as Moobot, an event’s organizers may decide to simply set up one
or more of these tools as the extent of their moderation setup. While I am unsure of how
common this is as a moderation strategy, it happens enough that one moderator I spoke to
said that they had previously tried to advise event organizers on their moderation setup, but
stopped because they were worried their advice would be taken as an insult.
Rarely, human moderators might be called upon to provide damage control for an event
while it is ongoing. I was an observer in one such instance, where a representative from an
event asked for help moderating their channel, and some volunteers responded. The speed
and surety with which the moderators started trying to implement their tools, bringing in
bots and manually taking action in chat, suggested this was not an unheard-of situation. The
moderator in question had even set their username to be bright yellow, in order to stand out
better in chat, so that whoever controlled the broadcasting account could spot them more
easily in the chat and therefore, grant them moderator status more easily.
Even with a human moderation team, present at the very beginning of the event, there
is no guarantee of training or adequate coverage of the event. Furthermore, not every orga-
nization will provide guidelines for their moderators, and even those with codes of conduct
or similar documents for their chats may not have them prominently displayed, once again
shifting the burden of both enforcement and education onto the moderators. The size of a
team may also cause issues. Somewhat counter-intuitively, though, my interviewees men-
tioned that an overly-large team could cause more problems than a small one, since a small

51
team could still moderate effectively given knowledge and familiarity with automated tools.
Suggestions of good numbers for an event moderation team ranged from “three or four” to
“twenty”, but those who advocated larger moderation teams also pointed out that this would
make the moderation team “as big as the broadcast team”, and therefore be somewhat im-
practical.
The issue they identified as hitting large moderation teams the hardest was commu-
nication. This was vividly displayed during DreamHack 2016, where a black professional
Hearthstone player, TerrenceM, was playing, and Twitch chat devolved into a constant stream
of racist commentary. Not only were the mods at the time unable to control the situation,
some allegedly joined in and offered to unban those users who had been participating in the
racist messages (Filewich, 2016). Less dramatically, though, large moderation teams that are
a result of organizers indiscriminately adding all applicants may open the door for deliber-
ate sabotage, as different groups of moderators conflict with one another. One moderator
recounted such a situation:

We have this…really big community of streamers. They have a monopoly on


everything but they don’t have a single decent moderator, but they still want to
have the power…because they have access to the [biggest event] channels them-
selves on Twitch, because they stream from them, they decide, “Oh look, it’s
these guys again, they’re moderating the chat but we want to have a monopoly.
Let’s demod them all.” Obviously we message Twitch…[Twitch] has to mod us
again, tell them not to do it again, they say, “Oh sorry, it’s not us, it was just
some new guy who wanted to do this, he’s just handling the camera and got
near the computer and decided to have fun.” And this happens at every event.

This is one example where having a large group of moderators does not lead to better
moderation. Such large groups may be a sign of indiscriminate hiring, heedless of exist-
ing social dynamics between groups of moderators with different styles and philosophies.
A larger group of moderators, lacking clear guidance from the event organizers or a single
respected head who can organize and enforce guidelines within the moderation team itself,
can be a recipe for disaster. This is part of why moderators create channels for communi-

52
cation amongst themselves, and one more experienced mod said “the thing I take pride in
having changed is the culture of communication within the moderation communities. Back
in the day there was very little to no communication at all…I tried to enforce always hav-
ing the mods online while they were moderating, that was actually one of the core rules we
moderators had. If you were moderating, please be in the group and be contactable and read
what I write in that group, if I messaged you please respond, because there was no way to
keep a moderation team cool if people were just uncontactable.” These attempts to create
clear lines of communication do not always extend beyond the moderation group, however.
Many of them said that contact with event representatives was rare, and if they could con-
tact a person, there was no guarantee that they would be responsive, or have access to critical
controls such as access to the broadcasting account. Some moderators also pointed out that
they could help with the event beyond dealing with Twitch chat.

A good example of that is [an event], two years ago. There was a kid up in the
front row with a sign that said, “They’re going A” and on the back, “They’re
going B”. So he would actually be flipping the sign and trying to cheat for his
team. That was one of the few times we actually got a hold of production almost
immediately. Security ran down and got the kid, took away his sign, almost
threw him out. And that’s one where—we need this all the time. There was one
other thing, it was too long ago and I can’t remember, some UK event. There
was some drunk dude that we could see when they panned the camera over to
the crowd that was just off his rocker, half-naked with a horse mask on. We were
like, “Uh, we can see this dude here we can see on stream, maybe call security?”

The emphasized section indicates that clear communication to people on the ground,
working at the event, is rare for moderation teams. A few moderators said that often, their
point of contact was the event’s social media manager. This was because the social media
manager, in charge of making live updates on the event’s accounts and creating the corre-
sponding graphics and short video clips, was also often watching the stream and staying
constantly updated. However, because different events have different organizational struc-
tures, that position or work may be done by one person, a few people, or no-one in particular,

53
further complicating communication between moderators and event organizers.
The reality of event moderation is not a clear hierarchy, with volunteer moderators below
event organizers. Rather, lines of communication are fraught. Volunteer moderators are at
once given considerable leeway with regards to judgement calls, yet little to no power to
communicate to other event staff. At once visible and invisible in chat, event moderators
fit uncomfortably into the hierarchies of event organizations even as they are clearly apart
from general spectators.

54
(a) Leftmost monitor.

(b) Middle monitor.

(c) Rightmost monitor.

Figure 3-4: An example of a moderator’s triple-monitor setup.

55
56
Chapter 4

The Social World of Moderation

4.1 The life of a moderator

How does a user make the jump to becoming a moderator? The majority of my intervie-
wees indicated that they started as moderators for a particular streamer’s channel, rather
than starting with event moderation. Through a combination of repeated and frequent vis-
its to a single stream, and a similarly sustained history of answering questions or providing
helpful criticism to the streamer or their moderators, became noticed and singled out as
particularly helpful regulars. One moderator got their start after pointing out problems in
a stream’s bot setup; another acquired a reputation as a helpful regular for CS:GO streams
by answering questions about event schedules, scores, and other basic information. This
method, of picking known community members, is consistent across all of the interviewed
moderators. At TwitchCon 2016, the moderator team for CohhCarnage, a popular streamer,
discussed the qualities they looked for when recruiting a new moderator.

“[It’s] how we choose our mods, it’s somebody who’s spent a lot of time in the
channel, a lot of time talking in the channel, and being, in our case, someone
with good vibes and a good countenance that is safe-for-work. Just following
our rules, being a member, and supporting the team…Mod selection is a bi-
ased process. Don’t think that it’s not. If people don’t know who you are, and
don’t know what you’re about, then you’re probably not going to make it to mod

57
status.”

The other, less-commonly discussed necessary condition for becoming a moderator is


the presence of ample free time. The three situations my interviewees highlighted were un-
employment, being a student, or being injured, thus freeing them up to invest an unusually
high amount of time in watching and participating in Twitch chat.
The jump from community moderator to event moderator is quite small. Some of the
interviewed moderators had previously been involved in setting up or running grassroots
esports tournaments, and therefore their involvement in Twitch event moderation was an
offshoot of their work for the tournament. Others were community moderators for esports-
adjacent streams, such as moderating for esports professionals, who were then invited to
work for a particular event. This movement was not all one-way, either, with some event
moderators taking on additional work as a community moderator after exposure to a team
or individual esports players after working esports events. Nevertheless, informal channels
of recruitment, mostly consisting of calls for moderators on private moderator-only groups
set up on platforms outside of Twitch or individual recommendations, seems to be common
and standard practice for event moderator recruitment.
Some larger, more established esports event moderation teams might have a more re-
cruitment process, involving written applications alongside informal recommendations. How-
ever, one head moderator admitted that the application was less important that simply ob-
serving newly-recruited moderators at work, saying that “on those applications they can
write whatever they want, but if they [don’t] do any moderation well then I won’t have them
on my team for longer than a week.” The use of applications can also create a lot of extra work
for head moderators, especially for popular events that might attract many eager applicants.
From a technical point of view, becoming a moderator is very simple. An existing mod-
erator or the broadcasting account can use a simple chat command, /mod username to
make someone a moderator; removing moderator status is similarly easy, /unmod username.
Moderators gain a small white-on-green sword badge next to their username to indicate
their status in chat. For this reason, moderator status is sometimes referred to as a “sword”
or “badge”.
A user with moderator status then has a different experience of Twitch chat. The most

58
apparent is that two buttons appear to the left of every user’s name in chat. These allow mod-
erators to ban or unban, and timeout for ten minutes, that user. Additionally, moderators
can set chat to different modes. These modes include slow mode, which forces users to wait
for a given amount of time between posting new messages; followers-only mode, which lim-
its chat to users that have followed the stream for a given amount of time; r9k-beta mode,
which only permits unique messages (and therefore cuts down on repeated or spammed
messages); and in extreme cases, subscriber-only mode, which only allows subscribed fol-
lowers to post to chat. This last one is considered a last-resort measure for event channels,
since there is little to no reason for a user to subscribe and pay a monthly fee for an event
channel, and consequently turning on sub-only mode for an event channel is like hitting the
off switch for chat.

What kind of training is provided to a new moderator? The prevailing method essen-
tially throws a newly-minted moderator into the deep end, relying on their natural trepi-
dation and hesitance to radically overhaul the existing system to avoid major issues. A few
moderators remembered a more hands-on mentorship, which affected their approach to-
wards new moderators. One said that “One of [the stream’s] more senior mods at the time
messaged me and set me up, gave me some advice and rules and stuff…And I try to do
the same thing with people who become mod, give them the same thing he gave me.” This
trepidation, or on the flipside, the confidence moderators feel for taking the initiative on
moderation action, versus following others’ leads, forms part of how moderators describe
their position within moderation hierarchies. Some of the moderators that I interviewed
considered themselves senior in some channels or for certain esports events, and junior in
others, precisely because in more unfamiliar channels they did not feel comfortable taking
a leading role in those moderation teams.

Because moderators, as discussed, tend to already be familiar with the moderation team,
they are aware (or are quickly informed) of the existing team’s hierarchy, generally from head
moderators telling them who to turn to in case of an emergency. The moderators at the top of
this hierarchy, whether officially titled head moderators or those wielding authority granted
through reputation and seniority, act as mentors to new moderators. New moderators are
expected to lurk, watch, and listen to more experienced moderators, and to follow in their

59
footsteps, though these expectations are rarely explictly stated. One interviewed moderator
remembered consciously modelling their judgement calls and actions on what other mod-
erators were doing, saying, “When I first got modded, I only banned the obvious, like people
who would drop the n-bomb. That was about all I did for a little while. Then I started look-
ing at things, like, ‘I would time that out.’ And about three, four seconds later, I’d see [the
head moderator] time that out. I did that for a couple weeks, then I was like, ‘I’ve got it,’
everything I wanted to do he was doing.”

Aside from being given moderator status, new moderators are granted access to moderator-
only discussion groups. These fast-paced chat rooms function as a combination training
ground, philosophical discussion space, and support group for moderators. While none of
the interviewed moderators spoke of any formal training program or systems, they did say
that these moderator-only spaces were where discussions about moderator policy happened.
“We pretty much talk about everything. It’s very open, and it’s not just the head moderators
or the chiefs or whatever their title is. We all discuss, all of the moderators, about policies,
ethics, how we should do things, why we do them, how we could be more effective,” re-
counted one moderator. This, combined with more experienced moderators checking on
newer moderators’ actions during events, forms the basic mentoring model for a event mod-
erator on Twitch. By being in these discussion groups, new moderators learn by watching
what more senior moderators consider good or bad actions, and get feedback as to which
of their actions were appropriate or not. In my observation, even experienced moderators
would frequently check in with the other moderators to ask for second opinions, such as
whether or not a user should be banned for having a slur in their username.

As newer moderators gain experience and confidence, they may take on more and more
responsibilities and seek out more knowledge related to moderation. The proliferation of
scripts, bots, and other automated tools means that basic programming knowledge is im-
portant for moderators to learn; in particular, knowledge of regular expressions is greatly
valued. Nearly all the moderators I spoke to talked about the importance of learning regex,
or at least understanding what it was and how it functioned. Moderators may also specialize,
“[incorporating] what they are studying, for example, into moderation…A lot of the up-and-
coming moderators last year have been young kids doing what they like, things they’ve been

60
doing in school.” Others read up on fields they believe may have use to them, “whatever you
can imagine, from economics to history to psychology.” This extra knowledge or experience
can be used to stake out a niche for themselves and their own moderation styles. This is one
way in which a newer moderator can make a name for themselves, and the desire to cre-
ate better tools for their work can be a powerful motivating factor for a moderator to teach
themselves how to program.

Finally, moderators who end up heading whole teams of moderators must take on man-
agerial duties, managing teams of people and liaising between an event organizer and their
team, taking over training and mentorship duties, in addition to performing moderation
work. Seniority and authority gained from increased responsibility also means that, in the
more informal world of the Twitch moderation community, one’s word gains weight in dis-
putes between moderators. In the words of one moderator, “I was one of the most, if not the
most, experienced person on the mod team. So if I just stated the result of an argument, that
was it, the discussion was usually over.” Part of their authority comes not only from their
history of work, but from the fact that moderators with a long career are able to remember
and bring up justifications and anecdotal evidence for setting moderation policy, further
lending credence to their arbitration. Senior, more experienced moderators’ work also ex-
tends beyond the technical and managerial, and into decisions involving moderation policy,
philosophy, ethical concerns, and pitching in with recruitment or training. Some larger, es-
tablished moderation teams might have training documents, pointing new moderators to-
wards certain tools and settings for them, as well as giving examples and best practices for
them to follow. These extend beyond decisions about content moderation. For example,
one extremely common recommended practice was turning on two-factor authentication,
a matter of online security. Again, all of these best practices documents are assembled from
moderators’ own knowledge and circulated within a relatively small closed group.

Given all of this, what might make a moderator leave? In a direct mirror to one of the
conditions for becoming a moderator, a lack of free time might lead to a slow decline in
the amount of work a moderator can do, which in turn may herald their departure from
the moderating community. Overwork, too, was cited as a cause for moderators leaving.
Recalling one high-profile departure, one moderator said that “he quit moderation…a year

61
ago, because he had had enough, because he pretty much overworked himself extensively
for free. And, you know, you can only do that for so long until your body caves in and you
just give up.” This was a rare acknowledgement of the lack of compensation for taxing mod-
eration work directly leading to a well-regarded moderator leaving. Even more striking is
its description of burnout, a spectre that hung over discussions about overwork and moder-
ator mental health in my interviews. While none of the interviewed moderators admitted to
feeling burnt out or even significantly affected by their work, all of them would admit that
they had heard of burnout or overwork contributing to someone else’s decision to leave.
Another is a disconnect between themselves and the communities to which they feel
they belong. This makes sense, given that many of the moderators I interviewed said they
identified as a fan of the esports scene before they became a moderator, and that the commu-
nity (whether of moderators or of that particular esport) was a reason for them to persist as a
moderator. One powerful anecdote came from a longtime Counter-Strike: Global Offensive
moderator, who recalled an incident that nearly made them stop moderating.

There were some kids on screen. There were eight or nine-year old kids on
screen…Chat said some fucking heinous shit. I’m sorry for my language, but
that’s tame compared to the things they said. I actually typed in chat, “I’m done,
I can’t do this any more.” Because chat just repulsed me…I slept 39 hours in the
past nine days and I was just too tired for it and I saw people saying, “I would
rape that kid,” or “Pedobear come here.” Just the worst shit.…that was the first
time I reached the level of disgust. It made the bile in my stomach churn.

The emotional and mental drain on moderators does not only come from viewing and
having to then remove the worst of Twitch chat. Every single moderator I interviewed had
said that they received harassment, abuse, and death threats; all of them told me this with
a laugh, signalling that it was so common an occurrence as to become routine and in some
cases, the basis for jokes. There are few other sources of support for moderators aside from
the aforementioned moderator-only chat groups, with some saying that they felt they could
not discuss their work because “it’s really hard to give a shit when, you know, it’s the In-
ternet…[people think] it doesn’t matter what people say.” In other words, the general view

62
that talk is cheap online is harmful for moderators trying to have their work taken seriously.
Even within moderator circles, all of them talked about the need to develop a thick skin or
to otherwise get used to constant abuse as either a matter of course for all moderators, or as
a requirement to persist in this role.

Moderators might also leave because of conflicts with other moderators, collectively
referred to as ‘drama’, and sufficiently dire disagreements may force even formerly well-
respected moderators to quit their teams. ‘Drama’ forcing moderators off of their teams was
something that hit some interviewees hard, especially if they were head moderators. One
head moderator, who could not work one event due to schedule conflicts, blamed himself
for not being there when a subsequent disagreement in his team lead to someone leaving.
“There was just that leading figure missing who could explain how and why the rules were
as they were, and as a matter of fact one person then left the moderation team. That was
slightly unfortunate, that was just because I didn’t put enough effort into keeping the mod-
eration team cool and controlled.” Instead of pointing at the harsh work conditions, lack of
support or compensation, or other structural factors that stress moderators, he attributed
the loss of this moderator to his failure to lead the team.1

Leaving the world of volunteer moderation for paid full-time work is also a fraught path.
It is currently vanishingly rare for a volunteer moderator to transition in to a paid modera-
tion job; at the 2017 TwitchCon panel, “Keeping the Peace through Moderation”, the pan-
elists could name “about five” moderators who managed to get a full-time paid position that
was directly relevant to Twitch moderation. The more common path, according to my in-
terviewees, was that volunteer moderation served as a way to network with people from the
esports scene, with the hopes that these connections would result in a job working for var-
ious event organizers. The resulting jobs were, anecdotally, most commonly in production
or social media management. While this makes some sense—Twitch moderation requires
one to be technically proficient in working with the Twitch platform, and with essentially
managing one’s own public reputation as well as upholding the event’s desired public brand

1
This is actually something of a recurrent theme: moderators seem well aware of the structural conditions
arrayed against them yet take personal responsibility for much of what goes wrong e.g. burnout, quitting,
disputes, or chat going nuclear. Unsure how to link it, but I feel this is a topic for elaboration in the ‘reality
ensures’ chapter.

63
or values—neither makes direct use of the expertise these moderators develop.

4.2 The relationship between moderators and Twitch chat

In order for moderation to function at all, a working relationship of trust must exist between
the moderators and the chats for which they moderate. The relationship between a moder-
ator and the chat which they oversee is not reduceable to a one-sided relationship of power,
or even a single stable mood. What is undeniable is that, for these large-scale event chats on
Twitch, there is constant friction and tension between these two groups, while the ghost of
other involved parties—event organizers, the personalities at the head of the most popular
Twitch streams, and Twitch staff—also influence the ongoing and dynamic relationship be-
tween mods and the chat. It is critical to remember that moderators are drawn from Twitch,
are embedded within esports communities, and often are moderating events for which they
are a fan. They understand the culture of Twitch, and, as a consequence of their work, have
had close experience with how it has evolved over the years, as well as a deep understanding
of how best to engage with chat. In other words, they make full use of sociability, “the col-
lective purpose of a community, the goals and roles of the individuals in a community, and
policies generated to shape social interaction” (Preece, 2000) to conduct their work.
In section 3.2.2, I previously outlined what red flags moderators look out for, and the
steps they take to address such messages. Understanding how these moderators view and
interact with Twitch chat is key to understanding the underlying logic behind these actions.
The interviewed moderators had a variety of metaphors with which they described chat,
the most vivid of which was describing it as “like a caged beast.” Recurring themes in all
these descriptions centered around its knee-jerk contrarian behaviour, extremely short at-
tention span, surprising cunning, and the contrast between individual chat users, as ex-
pressed through private messages to individual moderators, versus chat as a whole. This last
point is important: while moderators might variously describe individuals as “kids”, “trolls”,
or more neutrally “users”, they seem more concerned with the mob entity made up by the
mass of users, which is what I have referred to as ‘the chat’. This understanding of chat shapes
the moderators’ actions as well as moderation policy more generally.

64
According to the moderators I talked to, the prevailing wisdom for managing Twitch
chat was to be relatively permissive and to allow chat to have its fun, spamming memes
and reacting, provided they did not start using offensive language, or having one piece of
spam predominate and take over chat. Continuing on from the “caged beast” metaphor,
one moderator explained chat’s behaviour by saying that “if you cage it in too small, it will
get really angry and try to find any attack vector, in order to be as offensive and hateful as
possible. But if you set the bounds quite loosely, they will tone down their tone as well.”
Often, the trade-off that moderators decide upon is to allow disruptive spam, knowing that
removing it would provoke an outsized reaction. One moderator volunteered an anecdote
explaining this mindset:

An example I can give is in Rocket League chat, which is the official world cham-
pionship channel by Psyonix. People have just been spamming ‘gg’.2 And they
just spam it all day long and it’s annoying as hell, completely ruins chat. But
if you were to ban that, the backlash you would get would be way above and
beyond what you would get as a chat-ruining experience right now. So you just
leave it and let them spam their ‘gg’ and are happy that it isn’t anything racist.

These moderators understand their role and work as guiding chat more than control-
ling chat. One said that “chat has the attention span of a goldfish. You don’t have to go
heavy-handed and blacklist every little thing and knock out every little spam, because in
five minutes they’ll have something else they want to spam.” Rather than trying to set strict
filters that catch every instance of impermissible speech, moderators rely on their knowl-
edge of the rhythms of Twitch chat in order to knock certain spammed messages out of
circulation, or to suppress it temporarily. Because of the way in which repetition and spam
drives the pace and patterns of chat, moderators are not just concerned about offensive chat
messages (although they do, of course, remove them). Rather, they are particularly focused
on offensive chat messages that have a memetic, viral, or otherwise transmittable quality to
them.
2
Meaning “good game”, sometimes used to politely thank an opponent for a match, but in this context
meant sarcastically to deride someone for an easy win.

65
When asked what the point of Twitch moderation was, almost all of my interviewees
said that they understood the point of moderation as keeping chat fun and ensuring viewers
enjoyed the event. The one moderator who did not say that this was the primary purpose of
moderation instead saw their job as part of ensuring streamers and channels kept to Twitch’s
Terms of Service, thus protecting them from possible removal from the platform, but they
still expressed a moderation philosophy strikingly similar to the others. Granted, part of this
may stem from the fact that the world of large-scale esports event moderation is quite small;
certainly, there were a few key members of this community that my interviewees greatly
respected and pointed to as the source of their own values, policies and philosophies re-
garding Twitch moderation. We can see that moderators want to promote good experiences
for viewers, partially as a goal in and of itself, but also because they are aware that this makes
chat far easier to regulate than a stricter, more impermissive style.
The other piece of this puzzle comes from the fact that Twitch chat is very aware that
they are being moderated, and are also aware of some of the tools and methods with which
moderators regulate chat. For example, Moobot is not designed solely for moderation, and
it is easy to set up for any channel; therefore, Twitch chat users are quite familiar with its
operation since any of them could set up the bot themselves to look at its default filters and
settings. This knowledge is not always expressed maliciously. A moderator pointed out that
knowledge of bot-based moderation is itself the basis for jokes for chat, saying,

They actually @ moobot, they’re like, ‘moobot you shouldn’t be blacklisting that’,
you can see them throw fake insults at it. They’ve become more aware of the
tools we use and how we set those tools. If we set slow mode on for more than
ten seconds they’ll be like, ‘oh slow mode!’

The anecdote also shows that chat, collectively, can be very astute when it comes to fig-
uring out how the moderators of the channel are operating. Again, as previously mentioned
in section 3.2.2, some bad-faith actors know enough about how bot filters work to try and
modify their messages to get around the filter. By harnessing the observations of many view-
ers at once, chat can figure out different settings and use this as a new point of focus. This
same astuteness when it comes to figuring out moderator actions is used by the moderators

66
in order to change chat norms. Without having to explicitly state rules for chat, or by re-
ferring users to guidelines or similar, moderators can influence chat norms simply through
enforcement. One moderator described the long process of teaching some sections of event
chat to get used to new chat guidelines.

The first time he [hypothetical user] gets timeouted, maybe he won’t understand
why. And he tries this again. By using a lot of different software, we see he’s
doing this a second time and we timeout him again for one day, or half a day.
And you can see that people evolve, on the small scale—we have a chat with
300,000 viewers, obviously we can’t check everyone. But I started checking,
ten random people, what they do, what they did, did they adapt. And you can
clearly see the pattern. They adapt…in 90% of the situations, they know you
can’t say this any more.

What this moderator banked on was the ability of chat to first notice this new enforce-
ment, and then to realize that this new boundary was being regularly enforced. Equally, the
moderators rely on the bevy of third-party tools at their disposal to see the extent to which
these shifts are taking hold.
Moderators must be aware of other considerations when it comes to anticipating flare-
ups or fights in chat. I have previously mentioned that moderators may choose to take extra
precautions when they know that minorities will be shown on screen, by adding extra filters,
pre-emptively knocking out certain spam patterns, and being ready to redirect chat towards
less harmful games. However, for these events, moderators also have to keep apace with
scandals, fights, or other points of disagreement within the scene specifically. More gener-
ally, moderators also have to be attuned to wider Twitch cultural shifts. During IEM Ka-
towice’s finals, one flagged situation involved insults and provocation between subscribers
to two popular Twitch streamers, DrDisRespect and Forsen. They were easily identified be-
cause they were using custom emotes that they acquired by becoming paying subscribers to
their respective streams, and using them to mock subscribers or followers of the other. The
moderators indicated that a fight between DrDisRespect and Forsen’s followers was quite
common, since both had such huge subscription bases. They also suggested that the lax

67
moderation found on popular Twitch channels was a major contributor to general prob-
lems of Twitch chat toxicity.
Just as these moderators understood and saw ‘the chat’ as an entity distinct from the
individuals that comprise it, Twitch chat thinks of moderators as ‘the mods’, a collective
entity that obscures individual moderator. Jokes or popular spam messages in Twitch chat
about that tension between a channel’s moderators, and a sufficiently-cunning chat, abound.
Some are so common they have been collected on different sites that help users spam these
messages even more easily. Along with these messages and their endless variants, was the
frequent use of the phrase ‘Nazi mod’, to refer to what chat users believed was overly heavy-
handed moderation.

Figure 4-1: Some examples of popular mod-related spam.

As we can see in figure 4-1, three are about testing or poking fun at the limited attention
span and efficiency of the moderators, while the last aims to impersonate a moderator as a
joke. During my experience moderating during IEM Katowice, many users were not aware
of how to tell a mod from a regular user, with some claiming to be a moderator on the basis
of having a Twitch Prime badge3 , or being able to send coloured messages. Others, on seeing
messages deleted in real time, would attribute these actions to other non-moderator users
3
This badge is available for anyone who pays for Amazon Prime.

68
in the channel. More commonly, though, they would openly discuss whether they thought
a human moderator was present in the room or if it was merely an automated filter; the
consensus seemed to be that it was a bot. This confusion, with human moderators’ actions
mistaken for bot actions, was frequent enough that in the moderator-only Discord server,
some mods would jokingly introduce themselves as ‘often confused for a robot’.

However, breaking their silence to participate in chat is not always a good choice for a
moderator. Even if they engage with chat, their status means that they sometimes voluntarily
bar themselves from participation. Moderators are wary of being drawn into public debates
or conflicts, even if they want to intervene, with one saying that “sometimes I feel like I want
to say something but I can’t because I’m a mod.” Indeed, when I was moderating, a few users
started to argue about why there were so few women in professional gaming; when I said
that it was unlikely that genetic factors were to blame for this disparity, one user started to
complain about “mod baiting”, in other words, saying that the moderators were intentionally
stirring up conflict. Instead of understanding my opinion as something that I held as an
individual, which chat seemed to approve of when it was strictly in the realm of discussing
games, my dissenting opinion was understood as automatically done in bad faith or meant
to entrap them.

In other words, making moderators visible in chat is a careful act that is constrained
by informal rules and norms. It is something moderators learn to navigate based on previ-
ous experience and knowledge of Twitch chat’s general relationship to moderators, and to
that specific chat’s disposition towards this particular moderation team. Outside of making
their individual presence known, these moderators were also careful to present the team as
a unified force. Discussing inter-mod conflict publicly was considered extremely damaging
to the moderation team as a whole. One well-known and well-watched event, Games Done
Quick, has its volunteer moderators sign a non-disclosure agreement in order to prevent
leaks; while an extreme example, it shows the reticence of moderators and organizers to
make the details of their decision-making public.

For the interviewed moderators, the ideal Twitch event chat was one where permissible
memes and spam proliferated, without a single spam or type of spam dominating, and where
users either did not notice chat moderation or did and paid no attention to it. They saw the

69
free creation, remixing, and repetition of memes as a key part of what made viewing esports
on Twitch so enjoyable, and so it was important to them to create an environment where
new memes could be made.
The worst case scenario was a chat in open revolt against moderation; that is, instead of
chat complimenting the video stream and reacting to it, chat is arranged against the mod-
erators, possibly making a game out of testing the limits of the moderators (e.g. boasting
about being kicked multiple times, repeating memes related to moderator abuse or “Nazi
mods”). Successful chat moderation on Twitch requires the tacit approval of a majority or,
at least, silent plurality of Twitch chat users; when users feel dissatisfied with the moderators,
the situation can easily escalate, requiring more heavy-handed tactics to control the rate of
messages coming through.

4.3 What is good moderation?

Being modded adds a little green sword-shaped badge next to one’s username in chat. This
serves to mark out mods and has become slang: to “get your sword” is to become a mod-
erator. There is social cachet to having that badge by one’s name, since it is also the closest
thing to an authority figure in Twitch chat; in community spaces this also acts as a mark
of recognition, since moderators there are so often promoted from regular, generally well-
liked members. Several moderators described this badge as something that was generally
coveted; the general interpretation, joking or otherwise, of mods as power-hungry or other-
wise able to throw their weight around no doubt adds to the perceived value of moderator
status.
My interviewees identified two primary types of ‘bad moderator’: vanity moderators,
or those who got moderator status with no intention of doing moderation work, and rogue
moderators, which are compromised or malicious moderators. They also highlighted ‘badge
hunters’, which I class as a subset of vanity moderator; these are users who seek moderator
status on as many channels as possible while not putting in a commensurate amount of work.
As of the time of writing, Twitch provides no other easy way for streamers to distinguish
users in chat other than granting them moderator status, and therefore a moderator’s sword-

70
badge. Therefore, streamers occasionally “sword” their friends, significant donors, or other
users who they feel deserve recognition, but are not expected to perform any of the work of
a moderator. However, this phenomenon is rare in event moderation since there is little to
no reason why an esports event organizer would need to distinguish chat users in this way.

4.3.1 Rogue moderators, online security, and handling threats

Security issues are one of many considerations for event moderation teams. If any modera-
tor’s account is not sufficiently secured, it might be compromised, and since being a modera-
tor is currently an all-or-nothing proposition, a compromised moderator account is capable
of significant harm. These rogue accounts can mass-ban or -unban users, destroy bot set-
tings, cause that moderator to be removed from the channel, and generally cause chaos.
Therefore, the adoption of security measures such as two-factor authentication has become
a de-facto standard for the large-scale event moderators I talked to. A few recalled being the
target of attempts to compromise their accounts. One moderator remembered an incident
where someone ‘’leaked [their Twitch account] plain-text password in a crowded chat I was
modding. I had no idea who they were, all they did was post my plain-text password, and I
had to frantically change every single one of my passwords. I was lucky, he got it wrong by
one character. But that same night, he did that to other mods that I knew. He would, for ex-
ample, get on their account and start banning everyone and get them unmodded.” Another
described a very sophisticated hacking attempt, done to advertise a website.

One of our moderators got hacked…He was specifically targeted, they wanted
to make ads [originating] from a moderator…Moderators can’t demod other
moderators, only a streamer can, so we couldn’t do anything. And we tried to
timeout his followers who pasted the same advertisements. But we couldn’t,
because they had a script that purged all of our timeouts from his account. So
it was really pretty creative, I don’t know if they reached their goal, but it was
thoughtful. It was planned ahead. But we reacted pretty fast, we contacted
Twitch and after three minutes they banned that account. And afterwards we
banned all the people who followed him. All of this took two or three minutes.

71
Responses to rogue moderation accounts are still relatively rudimentary. Tools such as
moderation logs or mass demodding scripts can help as an emergency response; one mod-
erator who recalled trying to deal with this problem before the existence of these tools said
that, “for a while there was no moderation logs on Twitch, meaning we didn’t know who
was banning whom, so if I saw a lot of bad timeouts or bans the only thing I could do was
write in the group, ”Hey everyone, I’m seeing a lot of bad timeouts, can we maybe change
that, who has been doing that, please stop that.” If they wouldn’t read the group there was
nothing I could do. If there were hacked accounts…the only thing we could do was unmod
everyone, and see when it stops. And that person was the person that got hacked.” Preven-
tion, in the form of taking online security measures, is the preferred method of dealing with
the threat of compromised accounts.

However, the moderators did not mention being taught to take online security measures,
whether relating to keeping their account secure or keeping their personal information pri-
vate. When I directly asked one moderator, he replied, “No, never actually talked to any
mods about that kind of thing. Never heard anyone bring it up either. The only reason I
ever really thought of it maybe was because we laugh about the death threats.” Even then,
adoption of security measures, even in the face of constant threats, is not always uniform.
The same moderator said, of using safety or privacy measures, “I hope other people do.
Personally I don’t care.”

It was not lack of knowledge that kept him, and other moderators, from taking these
measures—when asked what should be done to ensure one’s online safety, he rattled off a
short list of things to do: ”Oh, you know…never put your real name on stuff, never have
your email account tied to something they could look you put for, never have your email
address public.” It was a cultivated sense of stoicism or apathy; as previously mentioned in
section 4.1, all the moderators I talked to were quick to say they were personally unaffected
by the many threats they received. In a notable point of comparison, they constantly justified
their reactions by pointing to the belief that others had it worse than they had. For example,
when asked about the abuse they had received, one moderator replied, “Oh yeah. So many
death threats!…Doesn’t bother me whatsoever…I’ve never really had anyone say, ‘I’m going
to find you, I’m going to stalk you,’ heavy stuff. I’m sure there are people out there who have

72
been doxxed or something like that, and do have a fear of that, a fear for safety, but I’m a
little unique in that sense.” Yet another moderator specifically brought up an instance where
someone attempted to threaten them by doxxing, to which they reacted with that same kind
of learned apathy.

I know there was one guy who that was trying to get my information, and he
went to this old website…that used to have leaked passwords and stuff. He got
my password off that site once and tried to track me and say, ‘I know your pass-
words and emails,’ and stuff like that…I actually just found it funny because it
didn’t really matter any more, and he said stuff like, ‘I have your address’…I
don’t really try to hide that either, it’s all public. I feel like I’m in a really safe
country so I don’t have to worry about anything happening, like getting swat-
ted. I don’t think any moderators have ever been swatted…he got it from a site
that posts leaked stuff that you have to pay for, so he paid money to get my in-
formation and then tried to track me with it, and I just found it funny that he
paid money for that. Because he tried to pretend that he was a hacker, but I
knew exactly where he was getting the information from.

It should be noted that, despite the stoicism demonstrated by the moderators I inter-
viewed, the moderation community frequently engaged in caring mental health and emo-
tional support behaviour for one another. This included a channel specifically for posting
cute animal pictures, and another for sharing memes or to joking about stressful aspects of
moderation. I also saw frequent occurrences of moderators checking in on one another,
talking each other through stressful situations or checklists of best practices for dealing with
harassment on Twitch. On this last subject, they were extremely knowledgeable: they de-
tailed specific actions that the user in question would have to do to build an actionable case
against their harassers for Twitch to become involved, as well as suggesting other steps to
protect themself while also reassuring the user in question that it was fine to step back from
moderation to engage in self-care. Some of the moderation guideline documents I saw also
included links for hopefully de-stressing content, such as a ‘cute animals’ link, in their ‘Re-
sources’ sections. Clearly, in private, moderators felt more comfortable expressing frustra-

73
tion, fear, and other negative emotions towards the work that they had to perform. In short,
despite the affected air of nonchalance that moderators display when asked about the stresses
of moderation, especially that which stems from the constant environment of abuse, they
are well aware of the toll it takes on those volunteers who perform this work.

4.3.2 Badge-hunters and proper moderation values

As for vanity moderators, the moderators I interviewed seemed to reserve special ire for
badge-hunters, or more generally, those users who seek moderator status with no intention
of putting in the work. Also called badge collectors, they were described as people who “ba-
sically just try to become a moderator on every chat they can get hold of, but they don’t do
anything.” This problem was not isolated to new moderators. Another moderator pointed
out that badge-hunters were sometimes already established and well-known in the moder-
ation community, saying that “we have a lot of moderators that are already known, but they
try to get as many swords as they can to just collect them for some reason, and you can see
they try to participate in the most events. But they don’t do anything. It seems to me like
they don’t care about their reputation.” Indeed, some praised the creation of moderation
log tools, which keep a record of every action taken by every moderator, as a way of seeing
which moderators were pulling their weight.
At the same time, a few of my interviewees were very critical of the perceived motiva-
tions behind users who volunteered for moderator positions; these individuals also tended
to be the ones who admitted to initially having the “wrong reason”, in their own words, for
becoming a moderator. These wrong reasons included believing that moderation was pri-
marily about “the power to remove anything from the chat that is not welcome”, or for the
more selfish reasons of “[getting] that extra popularity in chat and [having] your messages
more noticed.” They were more united on their descriptions of good moderators, and good
moderation motivations: wanting to help out the channel, streamer, or event, advancing the
interests of the broadcaster, and ensuring that chat-goers enjoyed themselves without that
enjoyment coming at the expense of either the event organizers or other chat users.
Interestingly, some of my interviewees maintained that this proper attitude towards mo-
tivation could not be taught. Simultaneously, others described a transition where “you

74
kind of learn to love the community as a mod instead of as a community member”, and
clearly, those moderators who admitted to seeking moderator status for selfish reasons have
changed their minds over the course of their career. While they were quick to talk about
changing minds on issues of policy, they did not raise as many examples of changing minds
on the issues of norms and values. This suggests to me that the proper or good moderation
values are being taught to new moderators, but through informal channels. One clear way
is the mentorship method by which these moderators learned to carry out their work: by
consciously modelling themselves on role models, it would not be surprising if they also
picked up their role models’ values. Others may have changed their mind after witnessing
moderator disagreements or discussions about particular moderation policies.

Indeed, some moderators talked about changing minds on the nuts and bolts of modera-
tion policy regarding permanent bans, both as a new moderator and as a mentor. However,
as they elaborated on the reasons behind why they thought the way they did about perma-
nent bans versus temporary timeouts, their reasons started to diverge: some believed it was
too heavy handed, some cited personal experience that it simply created more work dealing
with unban requests in the long run, others talked about mentors who explained their own
values and beliefs to them. For example, one moderator remembered being taught not to
hand out many permanent bans, saying, “I got taught when I joined moderation that we
don’t ban people…I am not afraid of banning, but I think that’s alright not to do, timeout is
enough. People tend to learn someday and if you ban them, they just create another account
and you have to ban them again, there’s no point.” In short, they clearly spoke about a change
in values that occurred alongside a change in policy, justified by personal and passed-down
precedents guiding anticipated outcomes and their estimated potential for future backlash.

This was most apparent over the contentious issue of emote bans. Twitch has long had
a reputation, and a problem, with toxicity around the use of global emotes depicting mi-
norities. Some of the more commonly used emotes, for these offensive purposes, are Tri-
Hard and cmonBruh, both depicting a black man; HotPokket, depicting a woman with dyed
blue hair; and Anele, a man in a turban. These global emotes are taken from well-known
Twitch streamers or Twitch staff, but have been appropriated and used for their surface rep-
resentation of these particular minority groups in order to express offensive sentiments, in

75
ways that are harder for filters to handle. The issue of offensive emote spam, particularly
spamming TriHard, came up during the 2016 DreamHack Hearthstone tournament, where
a black professional Hearthstone player was on-screen and Twitch chat responded with a tor-
rent of racial abuse, including spamming TriHard. More recently, professional Overwatch
player Félix “xQc” Lengyel was released by his team after an incident where he spammed
“TriHard 7” when reporter Malik Forte was on-screen.

Figure 4-2: Global Twitch face emotes often used in offensive messages. From left to right:
TriHard, cmonBruh, HotPokket, Anele

Multiple moderators told me that the norm a few years ago was to blanket-ban all use
of specific emotes that were used for offensive purposes. They were also very emphatic in
telling me that this was the wrong choice. From my observations of moderator-only spaces,
this is still an ongoing debate, but many of the more respected moderators now err on the
side of not blanket-banning emotes. In the words of one moderator,

There were some moderators who thought that auto-timeouting TriHard, just
TriHard without any context, was a good idea. And it was a ‘good idea’ for a
pretty long time…Now we have ohbot that checks the context. But the general
rule should be that any Twitch global emote should be allowed if you’re not
using it in a bad way. It’s always about the context, it always should be.

The various rationales given for context-checking and allowing global emote usage, even
if it made cutting down on racist or offensive spam harder, included the argument that global
emotes should by default be permitted, and that it was not a solution to remove emotes
depicting minorities from usage, given that the majority of face emotes are of white men.
However, this decision to check context, which can be automated because of the availability
of regular expression filters, is also in keeping with these moderators’ general belief that
permanent bans are last-resort tools; in short, instead of suppressing usage of these emotes,
they wanted to discourage their use in specific instances. In the words of one moderator,
moderation was “more about giving [them] the chance to say nice things instead of bad

76
things.” The decision made by these moderators to use more complex context-checking
filters, which screen what is being said before and after the emote itself, should be understood
as a policy decision made from repeated observation of the effect of blanket-ban policies, in
keeping with existing moderation philosophies, and implemented in automated tools as a
result of the affordances of said tools.
Because few esports organizations or game developers give their moderators chat guide-
lines, codes of conduct, or other guiding documents, moderators have had to develop stan-
dards for chat on their own. Even as they acknowledge that they are often the sole arbiters
of acceptable behavior in chat, many of the expressed discomfort with their position. When
asked if these organizations should be leaving these judgements to moderators, one replied
quite forcefully, “No! I would not trust the majority of people.” Others spoke of their at-
tempts to remain “impartial”, and of the responsibility they felt that they had to model good
behavior. Another directly told me, “You also have to understand, for me it’s really impor-
tant, that Twitch chat moderators are not—they can be wrong…I think you should look at
this too in your studies because it’s important how people make mistakes. In terms of how
they moderate or why they do this. Some people are power hungry, some people want to
show off how they do bad things.” Moderators also are aware of the ethical dimensions and
implications of their work. In addition to the reticence I described regarding their positions
as rule-makers, one moderator I talked to also expressed wariness over the ways in which
third-party tools, especially automation, could be deployed.

Well it’s become easier [moderating], but the problem also is, where do we cross
the line with how easy we can make it?…It’s also about how far can you take it.
There’s a fine line between, how much information should we keep about these
people in the chat? If a guy is being an absolute asshole in one chat, should we
ban him from another chat? …[A hypothetical banlist] would get abused, no
doubt about it, immediately I would guess. Somebody doesn’t like your opin-
ion? Well I have the power to ban you from thirty chats. You’re not allowed to
be here any more. That’s why we actually talk a lot of ethics nowadays and ways
to do things better. It’s so easy to go too far as well, I feel. A lot of moderators
are like, they want to have full control, they want to moderate everything, but I

77
feel like that’s not the correct way to go.

Ethical considerations also factor into moderators’ interpretation and handling of new
platform policies. That is, when Twitch announces new guidelines for streamers, moderators
are quick to adapt, since they know they will now be responsible for enforcing them even
if they may not be directly monitored. This is especially interesting in light of the fact that
not all moderators agree with all of Twitch’s Community Guidelines. In March 2018, Twitch
clarified their community guidelines by saying that “as a streamer, you are responsible for
the content on your stream.” As the community of moderators was quick to note, this could
easily mean that streamers could be held responsible for the actions of their chat if they
automatically displayed or responded to messages from said chat, for example if they had
an automated system that allowed donors to display any message of their choice or if their
stream included a live view of Twitch chat. A discussion broke out, with some moderators
arguing that streamers should not be held responsible for the actions of their viewers, while
others argued that, due to their influence and reputation they held, streamers should be
held accountable for how they shaped their chat. What is notable about this is the fact that
this discussion centered less on what moderators would have to do, but rather on how the
relationship between streamers, their moderators, and their chat should be configured. The
moderators of this community were able to comment and theorize, as well, on the historical
developments of Twitch chat that might have led to the current state of Twitch chat culture,
in order to justify or strengthen their arguments.

78
Chapter 5

Moderation Futures

In the scramble to comply with pressure on platforms to be seen regulating, who is left to
actually carry out enforcement? As our desire for greater and more visible moderation rises,
there is no guarantee that our understanding of the work of moderation will grow alongside
it. We began by noting that the pace of scandals around platform moderation seems to be
picking up. How might the future of moderation develop, at least for this group of mod-
erators? And what can we learn from this case study, especially with regards to what we as
online users, citizens, and participants ought to be asking?
I began this project with the intention of making clear the complexity of moderation
work, and in particular its cultural, communicative and social aspects. Yet at the same time
as I acknowledge that there are issues of scale at play for the largest social media platforms
that means it is hard to directly translate the volunteer model to moderation for them. How-
ever, I still believe that understanding the work done by volunteer moderators for their com-
munities, both communities of users and communities of moderators, is important in de-
veloping future support for them as well as considering what design for community growth
might entail.

5.1 Twitch, esports, and event moderation

As the esports scene changes, so will Twitch and its livestreaming competitors; the desire to
draw new audiences and sponsors is pulling at the scene and expectations of what is appro-

79
priate within it. Yet considerations of moderation rarely appear in popular discourse sur-
rounding it. Moderation only tangentially enters the discussion when it is centered around
the racist, toxic behaviour of both esports pros and of Twitch chat, and generally framed
in the context of how the organizations involved—game developers and esports teams—
choose to handle the situation. Platform exclusivity deals, such as ESL’s exclusive streaming
deal with Facebook for some of their largest DOTA 2 and CS:GO tournaments, further com-
plicate the future development of the esports moderation landscape. There is no guarantee
that these platforms have the kinds of moderation tools and affordances that allow volun-
teer moderators to carry out their work. Indeed, in the case of Facebook, their real-name
policy and policies against allowing multiple accounts means that moderators cannot rely
on anonymity to provide some measure of safety against reprisals.

At the same time, competitors to Twitch are pushing more transparent codes of conduct,
most notably Microsoft’s Mixer. Yet even when platforms champion transparent or easy-to-
understand policy, I do not see any of them champion moderation features, unless it is to
highlight automated chat message removal tools. Given all that moderators do, and the spe-
cific tools that allow them to do better work that have nothing to do with automated message
removal, the fact that this seems to be the only aspect of moderation that platforms are will-
ing to work on displays a deep gap between volunteer moderators’ practical experiences and
what developers are ready to give them as tools.

Yet, it is important to keep in mind the potential power of a moderating community.


Specifically, the communication networks and relationships formed between these volun-
teer mods has led to the formation of an organized force of workers, despite their lack of
compensation. i believe that the value of moderation work, as chronically undervalued as
it is, is nonetheless recognized as important to the formation of lasting communities, no
matter how dimly. What is concerning is the ongoing lack of attention paid to this group of
expert practitioners. Why do we see so little done to innovate for moderators, except from
other moderators?

80
5.2 Transparency, accountability, and reporting

Calls for transparency and accountability for platforms around their moderating decisions
are becoming increasingly common, as scandals about moderation, or the lack thereof, keep
cropping up. However, demanding increased transparency is not enough. All else remaining
equal, when we demand transparency for invisible work performed by invisible workers, we
wind up seeing nothing at all. Who do we want to be more transparent? About what? And
for whom? It is insufficient to demand knowing what is being removed and why. Instead,
we need to start asking about the philosophies of moderation that a given platform holds. All
platforms moderate, even, and especially, those that insist that they are neutral (Gillespie,
2018). While it is important that we know what is being removed, calls for transparency
need to encompass more.

As my interviewees have shown, continued enforcement is more than merely that. Our
focus on what gets removed exposes an appetite for understanding moderation only as it
pertains to the most visible aspects. Equally, it means we will always remain mired in a kind
of naive fascination with individual points when what we need to pay attention to is the tra-
jectory of moderation. In other words, we need to pay attention to the ways in which policy
is enforced, and how this enforcement creates and contributes to cultural and social changes
on the platform itself. We also need to understand that communities hosted on platforms
respond differently and nimbly to policy changes enacted by platforms. Finally, we need
to understand who and what operates the mechanisms of moderation in any given space,
and the conditions under which they labor. That is to say, we must expand our question to
include what that moderation is doing, and how it is being done.

I would argue that we, as users, deserve to see more than just notifications of content
removal. We need to have a moderation trajectory from the platforms on which so much of
our online social interactions occur. A moderation trajectory would include a philosophy
or some kind of articulation of what that platform believes to be good or proper modera-
tion, how it ought to be achieved, and a roadmap of broader goals, both policy-based and
concrete objectives, that can be achieved. A trajectory of moderation would be a statement
of moderation’s purpose with respect to the socio-technical context of the platform.

81
There are a few important implications of demanding moderation trajectories as a part
of the push for transparent platform moderation. Firstly, and most obviously, in order to
tell us what their moderation trajectory is, platform operators have to also know what it
is. This requires forethought. It means they must approach moderation proactively rather
than retroactively, to have had a plan in place before public outcry or evidence of missteps,
wrongdoing, or scandal.1
Secondly, a trajectory—a projected future course—requires more than platitudes. Hav-
ing a trajectory allows us to understand the moderation decisions of a given platform op-
erator in the context of past precedent and future goals., It means giving users the ability
to make sense of what actions and policies were present, are currently implemented, and
how they may change in the future. Again, it emphasizes moderation as a proactive series
of decisions with social repercussions beyond the immediate consequences of removal. This
is important because it will also require that we, the public, be able to fit both ongoing en-
forcement and points of failure into this trajectory. To criticize failures of moderation when
we are unsure of what we want, and what is being offered, as a moderation trajectory is one
thing; it will be another to criticize a sub-par trajectory of moderation.
The trajectory would make visible the work; next, we must make visible the human re-
alities of moderation. The workers are a fundamental and inseparable part of online mod-
eration. With respect to volunteer moderation, labour conditions are sometimes reduced
to the presence or absence of compensation, but it does not tell the whole story. Volunteer
moderators can be a well-networked and well-organized labour force. The absence of formal
support for volunteer moderators should not be confused with there being no support; the
lack of visible organization does not mean there is no organization.
If a given platform is going to offload any of the burdens of moderation onto its users,
then we must see that what they allow in terms of self-governance and community orga-
nization and moderation, as an integral part of its moderation trajectory. If they expect

1
Demanding a transparent moderation trajectory means thinking of moderation as going somewhere. I
purposefully choose this term over a more common descriptor, such as a moderation ‘strategy’, in large part
because I do not want to emphasize militarily-minded metaphor: there is not always an ‘us’ against a ‘them’, and
in any case the makeup of these groups are constantly in flux. I believe ‘trajectory’ is a better term because it
more strongly emphasizes the fact that there is a path to be laid, which has been neglected in favour of thinking
about reactive action. Granted, it implies a single or a clear path; nevertheless I think it is the better term.

82
self-governance, what has been provided, both in the design of the platform, and in the ex-
pressed values and norms that the platform expects these moderators to uphold? The two
are not mutually exclusive, and in order to understand and critique moderation we must be
conversant in both.
In an ideal world, we would have proper support for moderators, professional or other-
wise. Support would not just be limited to compensation: it would include the recognition
of moderators as an important constituency of a given space, and both the desire and ability
to take their voices seriously and to provide the conditions, material and otherwise, to allow
them to do their work to the best of their abilities. Platforms would be invested in publi-
cizing cogent, coherent plans of moderation, with track-records to bear them out, and be
able to elaborate on the philosophies guiding said moderation actions. As users, we would
be aware of and be able to participate in the work of moderation: not just the labour of flag-
ging, reporting, and so on, but the ability to decide on governance structures and access to
the information required to make informed decisions.
Realistically, though, I cannot expect even some plurality of users to bring themselves
to care so strongly about online moderation. People go on Twitch to watch livestreaming,
not to watch other people watching. But this is precisely why it is so important to recognize
volunteer moderators as a distinct group of invested users, occupied with organizing, tech-
nical, communicative, and social work. So long as moderation remains invisible, the actions
of a few will continue to have an outsized impact on our online lives. I am not trying to say
that volunteer moderators are either a force for good or a userbase to be feared. Rather,
they are motivated people who have found ways to take some small measure of control back
over their online lives, and of the communities to which they are tied. The human costs of
the work are exacerbated by the situation in which they, and we, are mired; not just costs
to the workers, but the costs of missteps, failures, and the subsequent need to be on guard
against malicious action. The structural factors arrayed against them are undoubtedly vast,
but not necessarily insurmountable. The work is hard, but worthwhile. And the first step is
to acknowledge what has already been done.

83
84
Bibliography

Braun, Joshua. 2013. Going Over the Top: Online Television Distribution as Sociotechnical
System: Online Television Distribution. Communication, Culture & Critique 6:432–458.
URL https://academic.oup.com/ccc/article/6/3/432-458/4054515.

Filewich, Carling. 2016. ”Enough is enough”: Confessions of a Twitch chat mod-


erator. https://www.gosugamers.net/hearthstone/features/39013-enough-is-enough-
confessions-of-a-twitch-chat-moderator.

Gasser, Urs, and Wolfgang Schulz. 2015. Governance of Online Intermediaries: Ob-
servations from a Series of National Case Studies. SSRN Electronic Journal URL
http://www.ssrn.com/abstract=2566364.

Geiger, R Stuart, and David Ribes. 2010. The Work of Sustaining Order in Wikipedia: The
Banning of a Vandal. CSCW 6:10.

Gillespie, Tarleton. 2018. Custodians of the Internet: Platforms, content moderation, and the
hidden decisions that shape social media. New Haven & London: Yale University Press,
1st edition.

Grimmelmann, James. 2015. The Virtues of Moderation. SSRN Scholarly Paper ID 2588493,
Social Science Research Network, Rochester, NY.

Humphreys, Sal. 2013. Predicting, securing and shaping the future: Mechanisms of gover-
nance in online social environments. International Journal of Media & Cultural Politics
9:247–258.

Jenkins, Henry. 2008. Convergence Culture. New York University Press, revised edition.

Jeong, Sarah. 2015. The Internet of Garbage. Forbes.

Kerr, Aphra, and John D. Kelleher. 2015. The Recruitment of Passion and
Community in the Service of Capital: Community Managers in the Digital
Games Industry. Critical Studies in Media Communication 32:177–192. URL
http://www.tandfonline.com/doi/full/10.1080/15295036.2015.1045005.

Kerr, Aphra, Stefano De Paoli, and Max Keatinge. 2011. Human and Non-human Aspects
of Governance and Regulation of MMOGs. 23.

85
Klonick, Kate. 2017. The New Governors: The People, Rules, and
Processes Governing Online Speech. Harvard Law Review URL
https://papers.ssrn.com/abstract=2937985.

Kou, Yubo, and Xinning Gui. 2017. The Rise and Fall of Moral La-
bor in an Online Game Community. 223–226. ACM Press. URL
http://dl.acm.org/citation.cfm?doid=3022198.3026312.

Kraut, Robert E., Paul Resnick, and Sara Kiesler. 2011. Building successful online communi-
ties: Evidence-based social design. MIT Press.

Latour, Bruno. 1992. ‘Where Are the Missing Masses? The Sociology of a Few Mundane
Artifacts’. In Shaping Technology/Building Society: Studies in Sociotechnical Change, ed.
Wiebe E. Bijker and John Law, 225–258. Cambridge, MA: MIT Press.

Massanari, Adrienne. 2015. #Gamergate and The Fappening: How Reddit’s algorithm, gov-
ernance, and culture support toxic technocultures. New Media & Society 19:329–346.
URL http://journals.sagepub.com/doi/10.1177/1461444815608807.

Matias, J Nathan. 2016. The Cost of Solidarity: A Quasi Experiment on The Effect of Joining
A Strike on Community Participation, in the 2015 reddit Blackout.

Niederer, Sabine, and José van Dijck. 2010. Wisdom of the crowd or technicity of con-
tent? Wikipedia as a sociotechnical system. New Media & Society 12:1368–1387. URL
http://journals.sagepub.com/doi/10.1177/1461444810365297.

Postigo, Hector. 2016. The socio-technical architecture of digital labor: Converting play into
YouTube money. new media & society 18:332–349.

Preece, Jenny. 2000. Online Communities - Designing Usability, Supporting Sociability. John
Wiley & Sons, Ltd.

Roberts, Sarah T. 2012. Behind the Screen: Commercial Content Moderation (CCM).

Shaw, Aaron, and Benjamin M. Hill. 2014. Laboratories of Oligarchy? How the Iron Law
Extends to Peer Production. Journal of Communication 64:215–238.

Silva, Leiser, Lakshmi Goel, and Elham Mousavidin. 2009. Exploring the dynamics of blog
communities: The case of MetaFilter. Information Systems Journal 19:55–81.

Terranova, Tiziana. 2000. Free labor: Producing culture for the digital economy. Social text
18:33–58.

Thomas, Bronwen, and Julia Round. 2016. Moderating readers


and reading online. Language and Literature 25:239–253. URL
http://journals.sagepub.com/doi/10.1177/0963947016652785.

Twitch. 2018. Community Guidelines FAQ Update. URL


https://blog.twitch.tv/community-guidelines-faq-update-a322c82b8038.

86

Centres d'intérêt liés